# Inside the Agentic Enterprise: How to Automate Your Organisation > Everyone runs the same models. What separates the 6% earning real value is redesigned workflows and the wiring between their systems. **Published:** 9 October 2026 **Last Updated:** 9 October 2026 **Source:** [https://www.elsewhen.com/reports/inside-the-agentic-enterprise/](https://www.elsewhen.com/reports/inside-the-agentic-enterprise/) --- **How to automate your organisation** For the people leading an agentic transformation. What an agentic enterprise is, why most AI programmes stall short of production and what the ones that work built instead. 16 minute read ## Redesign the work, then let agents run it 1. **Intelligence is a commodity.** Every competitor rents the same frontier models at the same price. 2. **Having it is not the same as gaining from it.** Nearly every large organisation now uses AI and very few earn real value from it. What sets those few apart is the wiring they built around the model. 3. **The wiring is the redesign.** Each workflow is written down step by step, with every step going to an agent, to plain code or to a person. One layer carries the work across your systems. So automating your organisation is a design and engineering job: deciding where agents act and where a human stays in the loop, then building the system that runs it. ## 00. Everyone has the magic box For a few pounds per million tokens you have the same frontier models as your fiercest competitor, improving at the same pace, on the same day they ship. Which means the differentiator has moved from the model doing the automating to what it automates: your organisation. **Every factory could buy electricity, and the winners were the ones that rewired around it.** Further reading: [Why the model is not your moat](/blog/why-the-model-is-not-your-moat/). Models converge and keep getting cheaper. The moat is the harness around them and the context it holds. The hard part is everything around the model: the business context the AI works in, the workflows it runs, the coordination of agents at scale, how your people teach it and the interfaces through which the work gets done. That layer, the one that joins a general capability to a particular business, is the only part nobody can sell you off the shelf. Deployment is a settled question. Gaining from it is not. In [McKinsey’s latest State of AI survey](https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai), nearly 90% of organisations use AI regularly, 37% can attribute any profit impact to it and 6% are earning material value. That 6% run the same models as everyone else. **What sets them apart is that they run AI as one programme for the whole business, held to the P&L from the start.** They decide where it will pay before they build, rebuild those workflows end to end along with the wiring that carries work between their systems and put as much into governance and their people as into the technology. So this report is about the wiring. How your operating model becomes software you own, and what has to be built before agents can run anything end to end. **Adoption is the easy part** Almost no pilot returns anything measurable, and close to half never reach production at all. The technology is already inside almost every large organisation. MIT put the cause at what it calls the learning gap: AI never integrated into real workflows, structures and context. - **95%** of enterprise generative AI pilots deliver no measurable return. ([MIT NANDA, 2025](https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo)) - **46%** of AI proofs of concept are abandoned before they reach production. ([S&P Global Market Intelligence, 2025](https://www.ciodive.com/news/AI-project-fail-data-SPGlobal/742590/)) - **78%** of the US labour force already works at a firm that has adopted AI. ([Federal Reserve Board, April 2026](https://www.federalreserve.gov/econres/notes/feds-notes/monitoring-ai-adoption-in-the-u-s-economy-20260403.html)) The AI Summit keynote: Inside the Agentic Enterprise [Watch the keynote](/blog/inside-the-agentic-enterprise/) ## 01. What is the agentic enterprise? An organisation where the work itself runs as software, with people directing it and making the decisions that carry weight. The workflows that make up the business are codified and run by agents, coordinated and governed in one place, while the people who know the business set the direction, apply judgement and stay in command of what matters. **What is an agent?** An agent is a model with tools (hands), memory (brain) and a goal (purpose), so it can act on your behalf and keep working alongside you. It runs a simple loop. Our previous report, [Building the Agentic Enterprise](/reports/building-the-agentic-enterprise/), covered the agents themselves; this one is about everything around them. Observe → Reason → Plan → Act → Remember → Repeat Today most enterprises look nothing like an agentic enterprise. They run hundreds of applications in silos, and the glue between them is manual: people rekeying data between systems, carrying context in their heads, moving each task to the next stage by hand. That manual glue is slow, hard to see into and easy to drop. **It is also where most of the time and cost go.** The agentic enterprise replaces that manual glue with an engineered one, so the work flows reliably and the people in it spend more of their day on judgement, relationships and decisions, and less on the mechanics in between. That is the operating model running as software, with people firmly in the loop where it matters. A chatbot bolted onto today's ways of working does not get you there. Further reading: [AI is more than a chatbot](/blog/ai-is-more-than-a-chatbot/). Why the chat window is the least interesting thing you can do with a frontier model, and what replaces it. ## 02. Why most of it never leaves the pilot The dominant way most organisations consume AI has a ceiling, and it is lower than it looks. Chat is often a poor interface for real work: it puts the entire burden of phrasing, context and follow-through on the person, one conversation at a time. And while putting a chatbot on every desk may feel like progress, the gain that matters arrives when the workflows themselves are redesigned around what agents can run. Meanwhile **tokenmaxxing**, pouring tokens across the organisation and treating consumption as a proxy for value, is proving to be a measure of activity and little else. Further reading: [The end of tokenmaxxing](/blog/the-end-of-tokenmaxxing/). What happens when boards start asking what all those tokens bought. The market has started to price this in. [Gartner](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027) counts roughly 130 genuine agentic AI vendors among the thousands claiming the label, and has a name for the rest: agent washing, the rebranding of assistants, RPA and chatbots as agents without the underlying capability. Buying activity and buying capability look identical on a slide. They diverge in production. Done right, the prize is the cost base itself, put to work. When agents run the workflows, operating expenditure stops being a fixed cost and becomes elastic capacity: the same spend produces more output each quarter as the system learns, and your people move up the value chain. Want to know which of your workflows is worth wiring first? [Book a working session](/contact/) ## 03. You have to redesign the enterprise to gain from new technology Having the technology has never been the same as gaining from it. The gains come from the redesign. History is blunt about this. Electric motors reached factories in the 1880s, but for 40 years they barely moved the needle. Plants that pulled out the steam engine, dropped in one giant electric motor and left the layout untouched saw almost nothing. Productivity exploded only in the 1920s, when factories were rebuilt around what electric power made possible: a small motor at each workstation, and machines arranged by the flow of work rather than crowded around a central driveshaft. Economists call the lag between a technology arriving and its payoff the productivity paradox, and the lesson holds every time: **the gains come from redesigning the work around the technology.** AI is at exactly this point. For most enterprises it has changed remarkably little, because they have bolted it onto workflows built for people. A copilot on a process designed for humans just runs the old process faster. The boom comes when the work itself is redesigned around agents. **The productivity paradox** Annual total factor productivity growth in US manufacturing. Electrification arrives in the 1880s. The payoff waits for the 1920s, when the factory floor is rearranged around what the technology makes possible. Total factor productivity growth in US manufacturing, 1880 to 2020, with the payoff arriving in the 1920s Sources: [David & Wright (1999)](https://www.nuffield.ox.ac.uk/economics/history/paper31/a4.pdf); [Field (2006)](https://www.cambridge.org/core/journals/journal-of-economic-history/article/technological-change-and-us-productivity-growth-in-the-interwar-years/99E3F7A9C33CA30278F06C1442A913E3); [US Bureau of Labor Statistics](https://www.bls.gov/productivity/tables/) ## 04. Every workflow is an algorithm If the gains come from redesigning the work, it pays to be precise about what the work is. All work is a series of steps to reach a goal, and every one of those series can now be codified. Onboarding a patient, approving an invoice, closing the books, migrating a portfolio. Each is a workflow: a trigger, a set of steps, a point where someone decides, an outcome written back. An organisation is just a collection of them, thousands, layered and interlocking. **That collection is your operating model.** For decades it lived in people's heads, in policy PDFs and in systems of record, because writing software to run it was slow and expensive. That constraint has gone. Coding agents now generate, test and ship software at a pace that was impossible 18 months ago, so any workflow can be codified and the operating model itself becomes something you build and own. Further reading: [Find your first agentic opportunity](/blog/finding-agentic-opportunities/). The workshop we run to pick which of your workflows to codify first, and the five tests we put each one through. Codifying a workflow means choosing who runs each step. **The human in the loop belongs wherever a mistake is costly, regulated or hard to undo**, and everywhere else the agent can act alone. **A company is a collection of algorithms** An enterprise is thousands of workflows, and every step in each one goes to an agent, to plain code or to a person. Redesigning the business means making that call step by step, across all of them. One workflow: Approving an invoice 1. An invoice arrives (trigger) 2. Read it and pull out the lines (agent) 3. Match it to the purchase order (code) 4. Explain any mismatch (agent) 5. Approve the exception (person) 6. Post it to the ledger (code) ## 05. The economics inverted Once any workflow can be written as software, the old reason to rent your software falls away. For 30 years the logic of enterprise software was simple. Building it was expensive and slow, so you rented it: a vendor amortised the cost of building once across thousands of customers and charged you per seat. **That bargain made sense while code was scarce.** Code is now cheap, and the market has repriced accordingly. Through the first quarter of 2026 roughly two trillion dollars came off software market capitalisations. The trigger was the arrival of agents that could carry sustained, multi-step knowledge work, which is precisely the work per-seat software was sold to support. For the first time on record, software's forward earnings multiple fell below the wider market. SaaS will survive, but the bargain has changed. Almost every incumbent now describes itself as agentic; strip the messaging and many are what they always were, databases that hold your data while you do the wiring. **That was a fair trade when software was scarce. It is a poor one now that code is cheap.** Further reading: [The case for leaving Salesforce, made by Salesforce](/blog/the-case-for-leaving-salesforce/). Benioff says the browser is optional and the API is the interface. Per-seat pricing does not survive that. **What the market pays for software, 2020 to 2026** Software's average forward price to earnings multiple falling from 84.1 times in 2020 to 2022 to 22.7 times in the first quarter of 2026, below the wider market for the first time. **$2tn** came off software market capitalisations in the same quarter, as investors repriced the sector. Software had never traded at a discount to the wider market before. The index has since recovered, but the premium for being software has not. Source: [SaaStr, average forward price-to-earnings multiple by period, March 2026](https://www.saastr.com/the-saas-rout-of-2026-is-even-worse-than-you-think-for-the-first-time-ever-software-now-trades-at-a-discount-to-the-sp-500/) ### Jevons paradox William Stanley Jevons set this out in [*The Coal Question*](https://www.econlib.org/library/YPDBooks/Jevons/jvnCQ.html) in 1865. When Watt made the steam engine more efficient, Britain burned more coal: factories built more engines and put them to entirely new uses, so total consumption rose. **Make a resource cheaper to use and you expand demand for it.** So when software gets cheaper, enterprises buy more of it. Worldwide IT spending reaches [$6.37 trillion in 2026, with software alone at $1.47 trillion](https://www.gartner.com/en/newsroom/press-releases/2026-07-27-gartner-forecasts-worldwide-it-spending-to-grow-14-point-2-percent-in-2026-totaling-6-point-37-trillion), up 15.5% on the year even as the market marked the sector down. **Enterprises are redirecting the software bill, from tools they rent towards capability they own.** Cheap software multiplies the role of software in the enterprise, and the organisations that gain are the ones that generate their own. **Cheaper to build, and more of it bought** Worldwide software spending, up 15.5% in a year, while the cost of producing it collapses. - 2025: $1.271tn - 2026: $1.468tn Source: [Gartner, Worldwide IT Spending Forecast, July 2026](https://www.gartner.com/en/newsroom/press-releases/2026-07-27-gartner-forecasts-worldwide-it-spending-to-grow-14-point-2-percent-in-2026-totaling-6-point-37-trillion) ### Once your operating model is software, it behaves like software What you generate is a description of how your business runs. A process that lives in a slide, a spreadsheet and four people’s heads cannot be versioned, tested or rolled back. Written down precisely enough for an agent to run it, the same process inherits every property software has had for 30 years, and the operating model turns from a deck into a deployable, composable asset. This is the part no incumbent vendor will say out loud, because the logic of how your business works is the one thing a competitor cannot copy and a vendor cannot sell you. **When the operating model itself becomes software, renting it from someone else is the last thing you want to do.** **Properties a workflow inherits once it is code** - **Version it** Every definition of a process, kept, dated and attributable. - **Diff it** See exactly what changed between two ways of working. - **Test it** Run a change against real cases before anyone lives with it. - **Roll it back** A bad change becomes a one-step revert. - **Fork it** Branch a variant for a region or a business unit, and keep both. > Governance stops being a document audited after the fact. *It becomes the rules the agents cannot break.* ## 06. What an end-to-end agentic solution looks like Whatever the use case, the same system sits underneath: the surfaces your people use, the agents that do the work, the coordination that keeps it honest, the context it reads, the systems it writes back to. A single agent on a laptop is a demo. An enterprise running thousands of agent executions a day across finance, compliance and customer service is an operation. The same shift happened with cloud computing: running one server is trivial, running a fleet is a discipline of its own. We call the reusable core under all of this the Productivity Platform. It is **built once, owned by you and reused across every use case**, so each new agent ships faster and cheaper than the last. It is model-agnostic, so as new models arrive they plug straight in and you are never tied to one vendor. Because you own the core, its value compounds with every use case. ### What the platform is made of - **The moat: Business Context** A structured model of how your organisation operates: connected data, explicit relationships, the rules that constrain decisions and the history that explains them. It grows with every interaction and stays yours. - **The workers: Headless Agents** Task-running agents with no interface of their own, callable from anywhere in your stack. They execute a defined process, use skills to follow it correctly and escalate when needed, with people in the loop where it matters. **They take the high-volume, repetitive work and clear the bottlenecks that throttle it.** - **The coordination layer: Orchestrator** **One agent is useful. Many agents, uncoordinated, is chaos.** The Orchestrator routes each task to the right agent, manages multi-step flows, shares context between agents and handles approvals and escalation. It also tracks outcomes so the system improves over time. - **The expertise layer: Skills Library** Domain expertise captured in a form an agent can load at runtime: how a job gets done, the rules, the guardrails. Anyone in the business can create and refine skills in natural language, **no coding required**. Expertise is captured once and reused across teams, and it stays in the hands of the people who hold it. - **The surface: Generative UI** **The interface is where adoption is won or lost.** Dashboards and short-lived apps produced on demand, shaped to the task at hand and delivered on whatever surface people already use. The interface comes to the work. ### The five parts, wired into your business **End-to-end custom AI · One holistic solution · Owned by you** - **CRM**: A system of record. It stays exactly where it is; agents read and write through governed connections. - **ERP**: A system of record. It stays exactly where it is; agents read and write through governed connections. - **CDP**: A system of record. It stays exactly where it is; agents read and write through governed connections. - **Contracts**: A system of record. It stays exactly where it is; agents read and write through governed connections. - **Content mgmt**: A system of record. It stays exactly where it is; agents read and write through governed connections. - **Agent**: The context agent. Reads your systems continuously and keeps Business Context current and trustworthy. - **Business context**: A structured model of how you operate: connected data, entity relationships, rules and governance. - **LLM**: The reasoning engine. Model-agnostic: we deploy the models you approve, on the cloud you run. - **Headless agents**: Digital workers that run whole workflows end to end. They take the repetitive volume; your people keep the judgement. - **Orchestrator**: The coordination layer. Routes tasks to the right agent, manages multi-step flows, handles approvals and escalation. - **Skills library**: Your experts teach agents how work gets done, in natural language. Captured once, reused everywhere. - **Generative UI**: Interfaces assembled around the task at the moment of need. The interface comes to the work. - **Web app**: A full application surface, for the jobs that need one. - **Dashboard**: Live oversight: what agents are doing, what they have done, and what needs a human decision. - **Chat**: Ask, instruct and approve in natural language, in the tools your people already use. - **User**: Your people. They set direction and own the decisions that matter, wherever the work needs judgement. - **3rd-party agent**: External agents plug into the same governed surface, built MCP-ready for what comes next. The AI Productivity Platform: the five building blocks in depth [Read the report](/reports/the-ai-productivity-platform/) ### Start with one workflow Nobody starts with all of it. You start with one of those algorithms running end to end, wired through the whole system, and you build only the parts that workflow touches. Prove that slice and the next one reuses most of it, which is where the compounding comes from. **Example workflows** #### Insurance (Underwriting operations) **Submission to quote**, Underwriting A broker submission lands as an email with attachments. The system reads it, pulls the risk data into your fields, checks it against appetite and returns a quotable file or a clear reason it cannot be quoted. 1. Read the submission and every attachment, whatever format it arrived in. 2. Pull the risk data into the fields your underwriters use. 3. Check it against appetite, capacity and referral rules. 4. Return a quotable file, or name exactly what is missing. 5. Send anything outside appetite to an underwriter with the reasoning attached. **Claim to decision**, Claims A claim arrives with documents attached. The system checks cover, pulls the facts out of the paperwork and either settles the straightforward ones or hands the adjuster a file that is already made up. 1. Match the claim to the policy and the cover in force. 2. Read the supporting documents and pull out the facts that decide it. 3. Settle low value claims that sit inside clear rules. 4. Hand everything else to an adjuster with the file already built. **Delegated authority review**, Compliance Coverholder bordereaux arrive monthly in whatever format the coverholder uses. The system normalises them, checks every risk against the binder and flags breaches while they still matter. 1. Take the bordereau in whatever shape it arrives. 2. Map it to your schema and reconcile it against the binder. 3. Check every risk for breaches. 4. Flag exceptions with the clause they breach. #### Freight and logistics (Pricing and RFQ desks) **Enquiry to rate**, Pricing desk A rate request arrives by email in the customer’s own words. The system reads the lanes, prices them against your tariffs and contracts and returns a quote the desk can send. 1. Read the request and pull out lanes, volumes and dates. 2. Price against live tariffs, contracts and surcharges. 3. Return a quote in your format, with the assumptions shown. 4. Escalate anything outside standard terms to the desk. **Tender to submission**, Bids A customer tender arrives as a spreadsheet with hundreds of lanes and a deadline. The system prices all of them, checks the result against capacity and builds the submission. 1. Read the tender pack and normalise the lane schedule. 2. Price every lane in the tender. 3. Check the result against capacity and margin floors. 4. Build the submission in the customer’s own template. **Exception to notification**, Operations Something has gone wrong in transit. The system spots it, works out who is affected and drafts the notification before the customer calls to ask. 1. Watch the status feeds for the events that matter. 2. Work out which shipments and which customers are affected. 3. Draft the notification with the revised plan attached. 4. Escalate anything carrying a penalty or a contractual clock. #### Lending and servicing (Migration and onboarding) **Portfolio to platform**, Migration A portfolio arrives from another servicer in their format with their conventions. The system maps it, reconciles it and lands it on your platform with the exceptions listed. 1. Take the seller file however it arrives. 2. Map every field to your platform’s schema. 3. Reconcile balances, arrears and history line by line. 4. List the exceptions a person has to decide. **Case file to review**, Compliance Regulatory review runs on a sample because a person can only read so many files. The system reads all of them and ranks what it finds. 1. Read every file in the population. 2. Check each one against the policy in force at the time. 3. Rank what it finds by customer harm. 4. Give the reviewer the evidence trail behind every flag. **Contact to action**, Servicing A borrower gets in touch. The system answers from your policies and their account, and stops short of anything that needs a person. 1. Take the contact from whichever channel it arrived on. 2. Resolve it against the account, the policy and the history. 3. Answer where the answer is unambiguous, and show the source. 4. Escalate vulnerability, complaints and arrears to a person. #### Professional services (Service desk, then finance) **Request to resolution**, Service desk IT, HR and finance requests all arrive in the same inbox. The system resolves the ones it can and routes the rest with the context already gathered. 1. Take the request from whichever channel it arrived on. 2. Resolve it against your own policies and systems. 3. Act where it is allowed to, and show what it did. 4. Route the rest with the context already gathered. **Enquiry to engagement**, Quote to cash A new enquiry has to be scoped, priced, conflict checked and papered before anyone does any billable work. 1. Scope the enquiry against your service catalogue. 2. Price it against the rate card and the client’s terms. 3. Run conflicts and onboarding checks in parallel. 4. Produce the engagement letter for review. **Redline to obligations**, Legal and commercial A contract comes back marked up. The system compares it to your position, flags what moved and pulls out what you have just signed up to do. 1. Compare the redline to your standard position. 2. Flag every departure and how far it goes. 3. Pull the obligations and dates into a list someone owns. 4. Escalate anything outside the negotiating mandate. #### Travel and attractions (Reservations and marketing operations) **Enquiry to quote**, Reservations A guest enquiry arrives in their own words with no dates confirmed and no structure. The system builds the itinerary, prices it and returns something a consultant can send. 1. Read the enquiry in the guest’s own words. 2. Build the itinerary against live availability and rates. 3. Return a priced quote the consultant can send or edit. 4. Escalate anything bespoke or high value to a consultant. **Search to booking**, Digital Guests search the way they talk. The system answers the question and returns something that can be booked. 1. Take the question in natural language. 2. Resolve it against live inventory. 3. Return results a guest can book there and then. 4. Read what converted back into the next answer. **Brief to campaign**, Marketing operations A campaign brief has to become dozens of variants, cleared by brand and legal, live in every market, with the results read back. 1. Take the brief and the audience definition. 2. Build the variants against brand and legal rules. 3. Route them for the approvals each market needs. 4. Push live and read performance back into the next brief. #### Public sector (Enquiry handling and casework) **Enquiry to answer**, Resident services A resident asks a question in their own words. The system answers it from your own material, with the source attached, and escalates the ones a machine should not be answering. 1. Take the question from whichever channel it arrived on. 2. Resolve it against your own policies, records and history. 3. Answer where the answer is unambiguous, and show what it was based on. 4. Escalate anything contested, statutory or high value to a person. **Request to disclosure**, Information governance An information request arrives with a statutory clock attached. The system finds what is in scope, drafts the response and holds it for an officer to sign off. 1. Log the request and start the statutory clock. 2. Find every record in scope across the systems that hold them. 3. Draft the response with exemptions flagged and the reasoning shown. 4. Hold it for an officer to sign off. Never auto release. **Application to decision**, Funding and casework An application arrives with evidence attached and criteria that have to be applied the same way every time. The system does the first pass and shows its reasoning. 1. Take the application and its supporting evidence. 2. Check it against eligibility criteria and the guidance in force. 3. Recommend a decision with the reasoning attached. 4. Send anything borderline or appealable to a case officer. The first step is picking the workflow where this would hit hardest. [Book a working session](/contact/) ## 07. How far do you go with agents? Autonomy works like a dial, and what it measures is how much the agent decides for itself. The right setting is a design decision, made workflow by workflow. Agents take on the repetitive coordination and the busywork between steps, so the people doing the work have more room for the parts that need them: deciding what is worth doing, shaping new products and improving the workflows themselves. People stay in command wherever judgement is the point of the step. In [PwC’s survey of US executives](https://www.pwc.com/us/en/tech-effect/ai-analytics/ai-agent-survey.html), 79% of companies have already put agents to work and two thirds report productivity gains from them, yet fewer than half are rethinking their operating model around what agents make possible. **The work is deciding which workflows earn a human in the loop, and what that person is checking.** **The autonomy dial** Most production systems we build sit across all three at once. - **Deterministic**: The agent follows a fixed blueprint. For steps that never vary and are expensive to get wrong: invoice to payment, where every run has to reconcile and be auditable afterwards. - **Human in the loop**: The agent does the work and a person approves at set points, in the tool they already use. For work where the exceptions carry the risk: contract review, where a lawyer signs off the clauses that matter. - **Fully autonomous**: The agent decides and acts with no checkpoint. For high-volume work where a wrong answer is cheap to undo: case triage, where a misrouted case costs a minute to move. Earned one workflow at a time. > Taste, judgement and accountability are the scarce resources, and *they stay firmly human.* ## 08. Put AI to work inside your business Custom software is now cheap to build and yours to keep. Work nobody would have automated two years ago suddenly pays for itself. The jobs coming to us are the ones where the money already sits: what the operation costs to run, what the licences cost to keep and the revenue these models make buildable for the first time. ### Automate what you do Agents inside the workflows that cost the most time and money. **£14m** of operating cost saved, every year In production · Financial services A global credit management company onboards every new loan portfolio in weeks rather than months, with every account checked instead of a sample. [Read the case study](/work/agents-onboard-loan-portfolios-in-weeks-not-months/) ### Replace what you rent The SaaS you rent, rebuilt as custom agentic systems you own outright. **£17m** of planner time saved a year In production · WPP Audience research that took weeks per brief now returns in seconds, on a system WPP owns outright, live in 62 countries. [Read the case study](/work/building-audiences-with-synthetic-data-in-seconds/) ### Grow what you sell Products and services that were not buildable before these models existed. **£50m** a year in additional booking revenue, based on reported financial results In production · Travel and hospitality Customers describe the trip they want and get real options back in under 250ms, matched on meaning against live inventory. [Read the case study](/work/ai-search-that-turned-browsers-into-bookers/) Not sure which of the three you are? Let’s talk. [Get in touch](/contact/) ## 09. A forward deployed squad One workflow, picked because it costs you the most, running on your data. Everything to this point is a set of design decisions: which workflows, how much autonomy, what the wiring looks like. What remains is who makes those decisions with you, and how fast they become running software. Agents carry around 80% of the execution. Forward deployed builders do the rest and sign off on everything that ships. You own what is built, IP included, and you see the measured return before committing further. - **Forward deployed: Inside your business** Our builders work in your environment, on your data, alongside the people who do the work. The people who scope it are the people who build it. - **Small team: Senior, embedded, sized to the problem** Product, data science and AI engineering, working alongside the agents that carry the execution. No pyramid, no handover. ## 10. What your competitors cannot copy The advantage now sits in the architecture you build around the model. That architecture is the context, the workflows, the coordination and the interfaces that turn a generic model into software that runs your business. So treat it as a programme for the whole business: decide where AI will pay, rebuild those workflows end to end, fund governance and your people alongside the technology and hold every workflow to the P&L. **Start with the one that costs you most, wire it through end to end, and each one after it reuses what the first one paid for.** > Everyone has the same magic box, so the enterprises that win the next decade will be *the ones that redesigned around it first.* ### Start with a conversation - Leon Gauhman, Co-Founder and Chief Product & Strategy Officer - Richard Henderson, Director of Business Development ## More from Elsewhen ### Inside the Agentic Enterprise Watch Elsewhen founder Leon Gauhman at the AI Summit, on why the model stopped being the differentiator and what has to be built around it before any of it pays. Leon Gauhman · Founder, Chief Product & Strategy Officer Watch: https://youtu.be/DHO6l71-E9g - [The AI Productivity Platform: Making AI Work at Scale](/reports/the-ai-productivity-platform/) - [Building the Agentic Enterprise: AI Agents & Multi-Agent Systems](/reports/building-the-agentic-enterprise/) - [Generative UI: From Generative AI to Adaptive, Agentic Interfaces](/reports/from-generative-ai-to-generative-ui/) ## Common questions ### What is an agentic enterprise? An organisation where the work itself runs as software. Agents carry whole workflows between systems while people set direction and make the decisions that carry weight, and the operating model is something the business owns. ### Why do most enterprise AI pilots fail? [MIT NANDA](https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo) found 95% of generative AI pilots deliver no measurable return, and put the cause at the learning gap: AI never integrated into real workflows, structures and context. The problem is the wiring around the model. ### What is the difference between AI automation and agentic workflow automation? Automation follows a fixed script and breaks when the input varies. An agentic workflow reads, reasons and drafts at the steps that need judgement, calls plain code where there is one right answer and routes to a person where the decision carries accountability. The coordination between those steps is the hard part. ### What is human in the loop, and which workflows need it? Human in the loop means an agent does the work and a person approves at set points, in the tool they already use. Autonomy works like a dial. Fixed, auditable steps can run deterministically, high-volume work where a wrong answer is cheap to undo can run autonomously and anything where the exceptions carry the risk keeps a person on the approval. ### What is the Productivity Platform? The reusable core every agent runs on: business context, headless agents, an orchestrator, a skills library and generative UI. It is built once, owned by the client and reused across every use case, so each new workflow ships faster and cheaper than the last. ### How long does it take to see a return? Elsewhen starts with one workflow running end to end and builds only the parts that workflow touches. The first one proves the return; the next reuses most of what the first paid for.