# Successfully implementing agents at enterprise scale

> Why agent pilots stall, and how an owned Applied AI Layer and five recurring agent patterns turn enterprise AI spend into a compounding asset.

**Published:** 17 November 2026

**Source:** [https://www.elsewhen.com/reports/implementing-agents-at-enterprise-scale/](https://www.elsewhen.com/reports/implementing-agents-at-enterprise-scale/)

---

Richard Henderson, Director of Business Development

November 2026

18 minute read

## Enterprises do not need more pilots.

To make agents work at enterprise scale, they need to establish and own an Applied AI Layer (a reasoning and action layer between their systems of record and their workflows) and populate it with a growing set of agents built to a few recurring patterns.

Concretely, the layer is software the enterprise owns, positioned between its systems of record (CRM, ERP and the document estate) and the workflows those systems feed. It reads the records, does the reasoning a workflow needs (interpreting, ranking, drafting, deciding) and hands a decision-ready output back at the point where a person commits. Every outcome becomes a new record the next decision draws on.

The layer means every agent works from the same enterprise context, under the same policies and through the same integrations, so each new workflow adds to what has already been built.

Even so, the task ahead is large: enterprise roadmaps run to dozens of functions, hundreds of workflows and thousands of agents. Reframing that estate around a few repeatable patterns lowers delivery cost and accelerates time to value.

> Build the layer iteratively, deliver the patterns consistently, and each new agent compounds on the last.

Two years of pilot activity have made the build-out of the layer and the adoption of the patterns the clearest predictor of whether agentic investment converts. This paper sets out how we help clients build the layer, deliver against the patterns and operate it in production, and what that changes about how value is delivered, governed and priced.

- **95%** of organisations in MIT NANDA’s research report no measurable return from their generative AI pilots
- **5** recurring patterns that every agent on the layer is built to
- **£50m** a year in additional booking revenue from one pattern in travel, based on reported financial results
- **$234bn** of enterprise application software spend at risk from agentic AI by 2030, on Gartner’s upper estimate

Sources: MIT NANDA, 2025; TUI Group and Elsewhen, 2026; Gartner, 2026. Each is linked where the figure appears.

## 01. The situation

Enterprise AI is in its second year of serious investment. The technology works, yet most pilots are caught in purgatory, neither cancelled nor scaled. Demos that work in isolation fail when they meet legacy systems. Teams define the same data differently. No one has set the rules for governance. And AI is added to old workflows instead of the work being redesigned around it.

**95%** of the organisations in MIT NANDA’s research reported no measurable return from their generative AI pilots

Source: [MIT NANDA, The GenAI Divide, July 2025, as reported by Fortune](https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/)

Across our own client base, a pilot that fails has failed on the operating model around it, meaning the architecture it runs on, the design of the work, the way it is delivered and the way it is funded. Companies rent copilot seats, embed vendors’ agents inside single stacks and stand up specialist tools inside single functions, then expect the sum to add up to structural transformation.

Within that base, the clearest predictor of success is how the problem was framed at the start. Clients succeed when they own their Applied AI Layer and build the agents that run on it from a few repeatable patterns.

## 02. The Applied AI Layer reframe

Autonomous agents decide for themselves how to reach a goal. With them, enterprises can reinvent workflows so that they run faster, cheaper, more accurately and at greater scale than the same workflows run by hand. What converts that potential into value is the Applied AI Layer, the reasoning and action layer that sits between the enterprise’s systems of record (CRM, ERP and the document estate) and the workflows they feed.

**How the layer changes the estate**

- **Pilots today**: Most enterprises run pilots, seats and point tools, each wired to a single system, with nothing shared between them or written back.

- **A layer between the records and the work**: What converts the potential of agents into value is an Applied AI Layer between the systems of record and the workflows they feed.

- **It reads, reasons and hands over**: The layer reads the records, reasons and hands a decision-ready output into a running workflow, at the point where a person commits. The agents doing the reasoning are each built to one of five recurring patterns.

- **Every outcome becomes a record**: Every outcome becomes a new record that the next decision draws on, so every decision leaves a trace the next one reads, each new pattern extends what the next one needs, and capability and returns grow together.

Source: Elsewhen. The pilots and workflows shown are examples.

**The data-quality decoy**

The legacy approach asks for clean data before work can begin, which often means a programme of several years that cleans data no workflow will ever use. The Applied AI Layer turns that around because it reasons over context where older systems followed fixed rules.

For each workflow, the agent makes sense of dirty, fragmented or duplicated records at the point of use, escalates what it cannot settle to a person and writes each resolution back to the system of record. The data is cleaned one workflow at a time, so clean data arrives as a by-product of the work and keeps improving as more workflows run.

The layer is the enterprise’s durable agentic asset. Its integrations are reused, its governance travels with every agent, and the enterprise can reason across a multi-vendor estate without trapping its intelligence inside one supplier’s software.

Four camps are competing to become the place your agents live, and the choice matters more than it first appears, because it decides who owns what you build.

The big software vendors are building agents into the systems you already run, such as CRM and ERP. It is the most straightforward path to adoption if your data already lives there, but your intelligence ends up locked inside one supplier’s product and you pay for it by the seat for as long as you use it.

The cloud and AI platforms (Google, Microsoft, Amazon and the model labs) let you use any AI model you like and swap between them freely. That openness is real, and the lock-in has simply moved up a level. The model stays swappable while everything that makes it useful, from the connection to your data and the memory to the controls and the way the agents are wired together, accumulates inside their platform, which leaves the model as the one part you can easily change.

The specialists sell finished tools for one job each, in legal work, customer service or clinical care. They work out of the box, and each function you automate means renting another one and stranding your data in another silo.

We take the fourth position. We build an enterprise-owned layer between your systems and your work that runs on whichever cloud and AI model you specify, and you own the value in it, from the connections and the logic to the controls and the accumulated work, none of which the other three positions leave with you. Each workflow delivered makes the base more robust, so the next one is delivered faster and at lower cost.

That is what the layer is for. It is one durable asset, and what runs on it is a growing set of agents, each built to one of a few recurring patterns.

## 03. Five recurring patterns

Across our implementations of the Applied AI Layer, five patterns recur, and we expect more to emerge as we take on more complex workflows. Each pattern has three parts: the reliable, deterministic systems it reads and acts through, the probabilistic judgement it makes and the point where a person signs off. The patterns differ in how those three combine, how they fail and recover, and how each extends the layer for the next.

**The five patterns**

The agent does the judging in all five. What changes is who sets the work going and who commits the outcome, and only one pattern lets the agent commit. Choose a pattern.

### Interpret and retrieve

Turns an imprecise human ask into a precise query the systems can answer.

- Person starts: A person asks, in their own words
- Agent judges: The agent works out what was meant
- Person commits: The person picks a result

How it fails: A wrong ranking or a buried answer

Risk and return: The lowest-risk place to start, and the fastest to deliver

This pattern works out what a person means from the way they ask. A customer can ask, in one sentence and with no filters or keywords, for a family beach holiday in Crete in August for under £3,000, and the agent surfaces the 10 options that match best. The customer picks the one they want, and the agent never acts on their behalf.

The request usually arrives vague, incomplete and missing the context a system would need. The layer infers the intent, maps it to a structured query and pulls the right answer from sources that do not share a common language, so nobody has to translate between the person and the systems. When it fails, the cost is a wrong ranking or a buried answer.

The commercial payoff is significant. TUI’s natural-language semantic search lifted search-to-book conversion by 28% among the UK visitors who used it, as [TUI reported](https://www.eqs-news.com/news/corporate-news/tui-delivered-resilient-q3-underlying-ebit-of-234-6m-at-cc-demonstrating-the-strength-of-its-business-despite-geopolitical-headwinds-and-one-offs-fy26-guidance-reaffirmed/179d539b-d12f-41af-a903-8192801cb022_en) with its results for the third quarter of the 2026 financial year, and that is worth £50m a year in additional booking revenue, based on reported financial results. Pointed at a catalogue of agents in place of a library of content, the same shape is what a conversational concierge does, reading each request and routing it to the right specialist agent. It is the lowest-risk starting point and the fastest to deliver.

### Interpret and recommend

Turns a machine output into ranked drivers and a next action, or a brief into a finished output.

- Output starts: A model’s output, or a brief
- Agent judges: The agent finds the drivers and the next action
- Person commits: The person decides and acts

How it fails: The wrong drivers, or an action outside the playbook

Risk and return: Ships with little risk and pays back well

A demand model forecasts that a site will underperform next quarter. On its own that number gives an executive nothing to act on, so the agent reads the model’s output and hands back the three drivers behind it in natural language, each with a specific action. The person reads, decides and acts, and the agent only recommends.

The agent works in both directions. It can read a machine output such as a forecast, a score or a model and turn it into ranked drivers, a clear explanation and a next best action, or it can read a brief written in natural language and turn it into a finished, structured output ready to use, such as a media plan, an audience definition or a proposal. Either way the person makes the call. It fails when the recommendation misfits, through the wrong drivers or an action outside the playbook.

The payoff shows in both directions. An international entertainment group’s site-level demand model now gives executives its drivers in natural language, mapped to specific marketing actions. At a global marketing group, audience planners turn a natural-language brief into a platform-ready audience definition, which returns £17m a year of planner time to the business and is in production across 62 countries. It is low risk to ship, and the return is high.

### Decide and act

The first pattern that acts, committing the enterprise to an outcome.

- Case starts: A case, such as a drawdown
- Agent judges: The agent decides, case by case
- Agent commits: The agent, inside limits people set

How it fails: Something irreversible left to the agent’s judgement

Risk and return: Takes the hardest design work for the largest gains

A drawdown needs approving before the funding window closes, with minutes to spare. The agent reads the position, checks it against the rules, prepares the transaction and sends it to a partner to sign. This is the first pattern that acts, committing the enterprise to an outcome, whether that means booking a trip, onboarding an acquired portfolio row by row or resolving a customer enquiry. The agent decides case by case, while people set the envelope of acceptable decisions and sign off at defined checkpoints, which spares them from approving every step.

The pattern works when the agent reasons, deterministic code executes anything irreversible under fixed rules, and people own the policy and the exceptions. It fails when something irreversible is left to the agent’s judgement and an action is taken that cannot be undone.

Within two and a half weeks, a major financial-services firm had a working prototype running in its own cloud environment, cutting loan-portfolio onboarding from months to weeks. This is the pattern with the hardest design work and the largest gains, because the work it takes on used to run for months.

### Synthesise and substantiate

Composes a document that has to stand up to scrutiny, and checks every claim against its source.

- Person starts: A person needs a document
- Agent judges: The agent drafts, then checks each claim
- Person commits: The person edits and signs

How it fails: A fabricated claim passing as a sourced one

Risk and return: Prevents the highest-liability failure and returns the most senior time

A diligence report, a legal memo or a clinical narrative has to stand up to scrutiny, and a single invented citation in one of them is a liability. The agent composes the document from the enterprise’s own library and from public sources, then checks each claim against its source and flags anything it cannot support before a person reads a word. The person edits and signs off.

It fails when a fabricated claim slips through dressed as a sourced one. The business case rests on reclaiming senior time, which is why we applied the pattern to standard-deck production at a large medical-communications firm. Harvey in legal work, Hebbia in investment analysis and Consensus in scientific research each sell a version of this pattern into a single vertical, and the layer runs it across the whole estate on a layer the enterprise owns. It carries the highest-liability failure to prevent and returns the most senior time.

### Monitor and escalate

Watches the whole stream of events and surfaces the few a person needs to see.

- Agent starts: The agent, watching events
- Agent judges: The agent decides what a person needs to see
- Person commits: A person acts on the escalation

How it fails: A missed signal, which fails silently

Risk and return: Always on, and the only pattern that fails by omission

Somewhere in the business, margin is being given away on one product line in one region, and nobody has noticed. The agent watches the whole stream of events, alerts, transactions and market signals and surfaces the few a person needs to see. Nobody asks this agent anything, because it starts the work itself, and the person receives an escalation and acts on it.

False alarms waste time, and the more dangerous failure is the alarm that never sounds, so the design guards against silence first. Calibrated well, the agent interrupts only when it matters. For a retail operator we are designing for, the pattern watches promotional mechanics and sell-through against forecast in a business where group margin runs in low single digits. It is always on, and it is the only pattern that fails by omission.

Source: Elsewhen implementations of the Applied AI Layer.

**What the first two patterns return**

- **28%** Uplift in search-to-book conversion among UK web visitors who use semantic search ([TUI Group, Q3 FY2026 results, 2026](https://www.eqs-news.com/news/corporate-news/tui-delivered-resilient-q3-underlying-ebit-of-234-6m-at-cc-demonstrating-the-strength-of-its-business-despite-geopolitical-headwinds-and-one-offs-fy26-guidance-reaffirmed/179d539b-d12f-41af-a903-8192801cb022_en))
- **£50m** A year in additional booking revenue, based on reported financial results ([Elsewhen client story, 2026](https://www.elsewhen.com/work/ai-search-that-turned-browsers-into-bookers/))
- **£17m** A year of planner time returned at a global marketing group ([Elsewhen client story, 2026](https://www.elsewhen.com/work/building-audiences-with-synthetic-data-in-seconds/))
- **62** Countries where the audience tool is in production ([Elsewhen client story, 2026](https://www.elsewhen.com/work/building-audiences-with-synthetic-data-in-seconds/))

**Actively guarding the loop**

When an agent acts, the risk moves to the person meant to check it, who can quietly defer to a machine that sounds sure. This is automation complacency. Decades of human-factors research in aviation and control rooms, reviewed in 2010 by [Raja Parasuraman and Dietrich Manzey](https://journals.sagepub.com/doi/10.1177/0018720810376055), shows it affects experts as much as novices, and that practice alone doesn’t cure it. In decide and act, the biggest risk is tired experts approving outputs that look right without really checking them.

Elsewhen designs against it in two ways. First, deterministic checks inside the layer clear the routine transactions and route only the low-confidence exceptions to a person, so fewer cases reach them and each one gets real attention. Second, we replace the static report with an interface the layer composes for the specific case, showing only the discrepancy, explaining the model’s reasoning and surfacing the historical precedent. A specific question draws a considered answer from the reviewer in a way a long document seldom does.

## 04. Inside the Applied AI Layer

At a major global professional-services firm, every practice group, every geography and every business function stood up its own pilots, each on a different vendor stack. That is pilot sprawl, paid for many times over.

The answer to sprawl is the Applied AI Layer, one holistic solution the enterprise owns and a single base that every workflow runs on in place of a new stack for every team. The five patterns describe how to build, and all five build on the same base.

Underneath every pattern sits a structured understanding of how the business works, covering its customers, costs, rules and history, which every agent draws on. Because it is built from how the enterprise actually operates, no vendor can replicate it, and it compounds with every workflow added.

The layer is less a pattern than the thing all five run on, and it is what makes each new agent cheaper than the last, because every workflow adds to it and the next one reuses what came before. It is an iterative programme that the enterprise runs and extends for itself.

**What the Applied AI Layer is made of**

Five parts, built up iteratively and reused across every use case, so each new agent ships faster and at lower cost than the last.

- **Business Context**: A structured model of how the organisation operates, covering its data, relationships, rules and governance. This first-party context is the part no vendor can replicate, and it compounds with every workflow.

- **Headless Agents**: Digital workers with no interface of their own that run whole workflows end to end, taking the high-volume, repetitive work so people keep the parts that need judgement.

- **Orchestrator**: Decides which agent runs when, manages multi-step flows, handles approvals and keeps people in command at the points that carry risk, so the automation always stays under control.

- **Skills Library**: The experts who hold the knowledge teach agents how the work is done, in natural language and without code, so the capability keeps improving without depending on Elsewhen.

- **Generative UI**: Interfaces assembled around each task at the moment of need, so the tool comes to the work and nobody has to learn another app.

Source: Elsewhen, [the Applied AI Layer](https://www.elsewhen.com/ai-productivity-platform/).

## 05. How the patterns compound

The five patterns compound because they share one Applied AI Layer. Every workload adds to it whether the enterprise plans for that or not, though only the deliberate builds compound. Returns compound with them: the second agent on the same layer costs less and lands faster than the first, and by the fifth the layer is doing most of the work, with the agent above it a thin overlay of pattern-specific reasoning.

The patterns are neither menu items nor a sequence. Whichever you build first extends the shared base for whichever comes next. Interpret and retrieve adds the entity resolution the others reuse, interpret and recommend adds the reasoning plumbing the others draw on, decide and act adds the audit trail the others sharpen, and synthesise and substantiate adds the source-checking the others call on. Monitor and escalate rides on the event streams the others produce. Order the builds by the business case.

Further returns come from what the layer lets the enterprise stop paying for, including legacy and data overhaul programmes that can be deferred or cancelled, SaaS tools the layer subsumes along with their per-seat licences, and the expensive agentic bolt-ons the systems-of-record vendors now push into CRM and ERP. Across our client base, avoided spend is often a material line in the business case, and in several cases it is the largest.

Pre-built vendor AI is a false economy, because its cost grows with adoption. The per-seat licence bill rises with every user, the data stays trapped inside the vendor’s ecosystem and every integration to another system needs a brittle custom connector, and when the subscription ends the intelligence goes with it. An owned layer asks for more investment at the start, and the cost of each agent falls as it grows, because it is built iteratively, with no speculative platform build, and runs against whichever models the enterprise chooses.

Over time the Applied AI Layer becomes one of the enterprise’s most valuable software assets, and it keeps paying back long after the individual agents that run on it have been replaced.

## 06. How to build

Our build approach breaks with what happens in traditional enterprise programmes, where delivery is organised by business function and the shared platforms meant to serve every function rarely earn reuse in practice. Delivering by pattern makes the pattern the reusable unit, sized between a shared platform and a single agent, and one squad owns it end to end.

The squad owns everything it needs to deliver the pattern, including the tooling, approvals, guardrails and delivery playbook, and it reuses what it has built for every agent that fits the pattern, whichever function that agent serves. Organising by function forces every department to rebuild the same pattern machinery for itself.

The benefits compound. Because the pattern’s machinery and risk approvals are established once, every later agent inherits them. Because the same squad keeps building inside the same pattern, it gets sharper and turnover drops. Sequencing becomes a design decision made on business priority, which takes the politics out of it, and governance is inherited, with the layer setting who can act and the pattern setting what they can do and how.

Fund each pattern as a whole. The first build in a pattern pays for the shared plumbing that every later build in the same pattern reuses, so later builds go faster and cost less, while funding each agent as its own build makes every one of them pay full price for that plumbing again.

**Who pays for the shared plumbing**

Each bar is one agent in the same pattern, serving a different function. When every agent is funded as its own build, each pays for the shared plumbing again. Funding the pattern means the first build pays for it and the other four inherit it.

Source: Elsewhen. Illustrative, and not to scale.

The Applied AI Layer is funded the same way, across the pattern builds instead of as a separate programme, with each build extending it and the next drawing on what the last one left. This is also how failure surfaces cheaply. If a pattern turns out not to fit the enterprise, that shows on the first build in the pattern, long before year two and seven one-off agents that never join up. Price accordingly, with the Applied AI Layer as the durable asset, the patterns as the units of delivery and the agents as the increments, and where outcomes can be measured, share the risk.

**The art of delivery**

Delivering by pattern takes judgement, and we make three calls. Sequencing starts with the pattern that has the clearest business case, whoever the loudest sponsor happens to be. Phasing sets up pattern squads that run in parallel and build agents one after another inside each pattern, each new one riding on the last. Composing assembles agents from different patterns, so a single customer resolution might monitor, retrieve, recommend, decide, act and synthesise, all on the same layer. Working delivery comes down to orchestration far more than to tooling.

## 07. The strategic imperative

Owning the layer is the better architecture, and the window to do it is closing. The vendor landgrab is running fast, with Microsoft, Salesforce, ServiceNow, Palantir and Google all competing to become the layer an enterprise’s agents run on.

**$234bn** of enterprise application software spend, at most, that Gartner estimates is at risk from agentic AI by 2030

Source: [Gartner, press release, 1 July 2026](https://www.gartner.com/en/newsroom/press-releases/2026-07-01-gartner-says-us-dollars-234-billion-in-enterprise-application-software-spend-is-at-risk-from-agentic-artificial-intelligence)

Every month an enterprise waits, more of its workflows are built on someone else’s stack, and they get harder to take back. The choice to own the layer has a window of the next 12 to 24 months, and an enterprise that does not make it deliberately will find it made one vendor integration at a time.

Most pilots die in purgatory because they were built as pilots, one at a time, on someone else’s stack and joined to nothing. The way out is to stop building pilots and start building the layer. Pick the pattern with the clearest business case, build it well and let each agent after it arrive faster and at lower cost on the same Applied AI Layer.

The models will keep changing, so the lasting value sits in the layer you own, which compounds and stays yours. Own the layer, and the next 12 months of work leave the enterprise with an asset it keeps.

## 08. Questions and answers

### What is a reasoning and action layer?

It is software an enterprise owns, positioned between its systems of record, such as CRM, ERP and the document estate, and the workflows those systems feed. It reads the records, does the reasoning a workflow needs, from interpreting and ranking to drafting and deciding, and hands a decision-ready output to the point where a person commits. Every outcome becomes a new record that the next decision draws on. The Applied AI Layer is the name Elsewhen gives to the one it builds with its clients.

### Why do most enterprise AI agent pilots stall?

The technology works, yet most pilots are caught in purgatory, neither cancelled nor scaled. Across Elsewhen’s client base the cause is the operating model around the pilot, meaning the architecture it runs on, the design of the work, the way it is delivered and the way it is funded. Pilots built one at a time on someone else’s stack, and joined to nothing, rarely convert.

### What are the five recurring agent patterns?

Interpret and retrieve works out what a person meant and pulls the right answer. Interpret and recommend turns a model’s output or a brief into drivers and a next action. Decide and act commits the enterprise to an outcome inside limits people set. Synthesise and substantiate composes a document and checks every claim against its source. Monitor and escalate watches events and surfaces the few a person needs to see.

### What is the difference between the layer, an agent and a pattern?

The layer is the durable apparatus an enterprise owns and builds on: the connections to its systems, the reasoning, the controls and the governance. An agent runs on the layer to do a specific class of work, reading, deciding, drafting or acting. A pattern is a recurring way of building an agent. The enterprise owns the layer, builds agents on it and builds each agent to a pattern.

### What is Elsewhen’s Applied AI Layer?

It is how Elsewhen builds the reasoning and action layer for its clients, from five parts that are built up iteratively and reused across every use case: Business Context, Headless Agents, an Orchestrator, a Skills Library and Generative UI. It runs on whichever cloud and AI model the enterprise specifies, and the enterprise owns it.

### Do you need clean data before deploying AI agents?

No. The layer reasons over context, so for each workflow the agent makes sense of dirty, fragmented or duplicated records at the point of use, escalates what it cannot settle to a person and writes each resolution back to the system of record. The data is cleaned one workflow at a time, as a by-product of the work.

### Should an enterprise use its software vendors’ built-in agents or own its layer?

Vendor agents are the most straightforward path if your data already lives in their systems, though the intelligence stays inside one supplier’s product and is paid for by the seat. Cloud platforms let you swap models while the connections, memory and controls build up on their side, and specialist tools add a silo for each function. An owned layer keeps the connections, the logic, the controls and the accumulated work with the enterprise, and each workflow makes the next one faster and cheaper to deliver.

### How do you stop people rubber-stamping an agent’s decisions?

Automation complacency appears in experts as readily as in novices, so the design has to carry the load. Deterministic checks inside the layer clear the routine transactions and send only the low-confidence exceptions to a person, so each one gets real attention, and the reviewer sees an interface composed for that case, showing the discrepancy, the model’s reasoning and the historical precedent.

### How should an enterprise fund and deliver agents?

By pattern. One squad owns each pattern end to end, with its tooling, approvals, guardrails and playbook, and reuses them for every agent that fits the pattern, whichever function it serves. The first build in a pattern pays for that shared plumbing and every later build inherits it, so later agents arrive faster and at lower cost. Where outcomes can be measured, share the risk.

## Keep reading

- [The AI Productivity Platform: making AI work at scale](https://www.elsewhen.com/reports/the-ai-productivity-platform/)

- [Why enterprise AI fails to deliver](https://www.elsewhen.com/blog/why-enterprise-ai-fails-to-deliver/)

- [AI search that turned browsers into bookers](https://www.elsewhen.com/work/ai-search-that-turned-browsers-into-bookers/)

### Start with a conversation

- Richard Henderson, Director of Business Development
- Leon Gauhman, Co-Founder and Chief Product & Strategy Officer