Skip to content
A bearded man in glasses talks with his hands as colleagues listen around a table in a sunlit studio with tall windows and plants

Successfully implementing agents at enterprise scale

18 minute read

Executive summary

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.

Section 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

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.

Section 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.

Figure 1

How the layer changes the estate

  1. Step 1

    Pilots today

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

  2. Step 2

    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.

  3. Step 3

    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.

  4. Step 4

    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 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.

Section 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.

Figure 2

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.

Pattern 1

Interpret and retrieve

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

  1. PersonStartsA person asks, in their own words
  2. AgentJudgesThe agent works out what was meant
  3. PersonCommitsThe 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
Read the full pattern

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 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.

Source: Elsewhen implementations of the Applied AI Layer.

Figure 3

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
£50m
A year in additional booking revenue, based on reported financial results
Elsewhen client story, 2026
£17m
A year of planner time returned at a global marketing group
Elsewhen client story, 2026
62
Countries where the audience tool is in production
Elsewhen client story, 2026

Section 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.

Figure 4

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.

  1. Step 1

    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.

  2. Step 2

    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.

  3. Step 3

    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.

  4. Step 4

    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.

  5. Step 5

    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.

Section 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.

Section 06

How to build

A woman in glasses points to something on the table as she talks, while two colleagues at laptops listen in a sunlit, plant-filled studio

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.

Figure 5

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.

Each agent funded as its own build

The pattern funded once

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.

Section 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

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.

Section 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.

Start with a conversation