The engagement
TheAgentBuild.
Most of the work in a business is a workflow: something arrives, someone reads it, checks it against what they know, does the next thing, and tells someone. AI can improve and automate it: an agent that runs it end to end, or one that automates the routine steps so your people spend their time on the judgement calls. How far it goes is your decision, made per workflow. The engineering is the same either way: connecting it to your systems, proving it on your real cases, and making it know exactly when to hand back to a person.
-
01
Inbound email and triage
A shared inbox where every message is read, matched to a customer, case or order, and either answered from your own knowledge, routed to the right person, or escalated — with the reasoning attached so whoever picks it up knows why.
-
02
Document intake and checking
Invoices, applications, contracts, forms and attachments arriving in every format. The agent extracts what matters, checks it against your rules and the records you already hold, files it where it belongs, and puts every exception in front of a person with the evidence.
-
03
Tickets and service requests
Requests that follow a known path — resets, status checks, changes of detail, first-line diagnosis, standard approvals — handled from open to close, with a full record of what was done and anything off the path handed to your team.
-
04
Reconciliation and matching
Two systems that should agree and never quite do: payments and invoices, orders and deliveries, bookings and rosters, records held in two places. The agent matches what it can, explains what it cannot, and never guesses.
-
05
Onboarding and provisioning
New customers, suppliers or staff, chased for the same information by the same three people. The agent collects it, verifies it, sets up the accounts and access, and keeps everyone told where things stand — with sign-off before anything goes live.
-
06
Reporting and monitoring
The weekly pack assembled from six spreadsheets and four people's memories. The agent gathers the figures from the source systems, drafts the commentary, flags what moved and why, and has it ready before anyone asks.
-
07
Scheduling and coordination
Appointments, deliveries, shifts, site visits and the back-and-forth that surrounds them. The agent proposes, confirms, reschedules and chases across email, calendar and your own booking system, within the rules you set.
-
08
Research and enrichment
Every record that needs looking up before someone can act on it — a company, a property, a claim, a candidate, a supplier. The agent gathers what is public, cites where it came from, and leaves a complete file rather than a starting point.
Your call
Runs it, or helps.
Not every workflow should run itself, and not every team wants it to. Some want the queue cleared overnight with only the exceptions waiting for them. Others want a colleague that reads, checks, drafts and suggests, so their people decide faster and keep the final say. Both are the same agent with the dial set differently, and the setting is yours to change.
- Runs it
- Works the queue, hands back exceptions
- Helps
- Prepares the work, your team decides
- In between
- Runs the routine, drafts the rest
- Set during
- Discovery, per workflow, changeable
88%
Use AI — but only 39% see EBIT impact
AI usage is widespread, but fewer than four in ten organisations report any impact on enterprise-level EBIT. The difference is not the tool. It is whether the workflow around it was actually changed.
A workflow, before and after.
The same shape appears in every business we walk into. The agent does not change what the workflow is for. It changes how much of it runs itself: all of it, or the routine steps with your team keeping the decisions.
Typical today
- Work arrives by email, portal, post and phone into one person's queue
- Someone reads each item and works out what it is and who it is for
- The same details are re-keyed from one system into another
- Checks depend on who is in that day and what they remember
- Simple cases wait behind hard ones
- Chasing missing information is a job in itself
- Status lives in someone's head, or a spreadsheet nobody trusts
- Nothing happens overnight, at weekends or during holidays
- Nobody can say how long anything actually takes
- The one person who knows the process cannot take leave
With an agent
- Every channel lands in one queue the agent works continuously
- Each item is classified, matched to its record and routed on arrival
- Data moves between systems once, through their own interfaces
- Every check runs every time, against the same written rules
- Known cases complete themselves; people get the ones that need them
- Missing information is requested, tracked and chased automatically
- Status is visible per item, with the agent's reasoning attached
- The queue is worked around the clock, within the limits you set
- Every step is timed and logged, so the process can be measured
- The process is written down, tested and no longer lives in one head
46%
Of AI proofs of concept are scrapped
The average organisation abandons almost half of its AI proofs of concept before production. A demo that never touches a live system is the most common way to spend a year on AI and change nothing.
What you get.
An agent is only finished when it is running on live work inside your systems, with your team able to see what it did and why. Everything below is what an engagement leaves behind.
| Stage | What we do | What you get | Who decides |
|---|---|---|---|
| Map | Walk the workflow with the people who do it, step by step | The process written down, with every rule and exception | You |
| Define done | Agree what a correct outcome is, how much the agent does on its own and where your team stays hands-on | A test set built from your own real cases | You |
| Connect | Integrate with the systems the workflow touches, read-only first | The agent seeing what your team sees | Us |
| Build | Build the agent and its hand-back rules against the test set | Every case passing, or explained | Us |
| Pilot | Run it on live work alongside your team, in shadow then in earnest | A side-by-side record of what it did and what they would have done | You |
| Hand over | Document, train, and give you the code and the dashboards | An agent you own and can see into | You |
| Operate | Watch it, re-test it, and adjust as the business changes | A monthly report on what it handled and what it handed back | Both |
The measure is not how much the agent does. It is how much it gets right — and whether a person heard about the rest.
An agent that quietly does the wrong thing at scale is worse than the manual process it replaced. So every agent we build is tested against an evaluation set drawn from your own real cases before it touches live work, and re-tested whenever the rules, the systems or the model underneath it change. When it is not sure, it stops and asks. That is a feature, and we design for it.
Guardrails
Bounded by design.
Every agent has a written scope: the systems it may read, the systems it may write to, the actions it may take on its own, and the ones that need a person. Those limits are enforced in code, not in a prompt, so they hold when the model is wrong.
Every action is logged with the reasoning behind it. Your team can see what the agent did, why, and what it was looking at when it decided — and can correct it, which the agent learns from at the next review.
We are not tied to a model, a cloud or a platform, and we take no vendor commission. The agent runs on whichever combination passes your evaluation at a cost that makes sense, and it is built so that combination can change without rebuilding the agent.
It does the known work. It hands back the unknown.
Map it.
Build it.
Run it.
Three stages, one workflow at a time. Start with discovery — if the workflow is not a good fit for an agent, we will say so and tell you what would be.
-
01
Workflow Discovery
Pick the workflow. Walk it with the people who do it, capture every rule and exception, define what a correct outcome is, and decide how much the agent does on its own and where your team stays hands-on. Ends with a written design and a test set of your real cases.
-
02
Agent Build
Build the agent into your systems, prove it against the test set, then pilot it on live work — in shadow beside your team first, then taking the queue. Handed over with the code, the dashboards and the documentation.
-
03
Operate
Monitor what it handles and what it hands back, re-run the evaluations as rules, systems and models change, and extend it to the next workflow when the first one is boring. Someone needs to be watching, and it can be us.
If someone does it by hand every day, start here.
You do not need an AI strategy, a data platform or a team of engineers. You need one workflow you can describe, and the systems it lives in.
- A queue that grows faster than the team working it
- The same information re-keyed between two or more systems
- A shared inbox that someone spends the morning triaging
- Checks that depend on who is in and what they remember
- A process only one person fully understands
- A team that wants help with the work, not to hand it over
- An AI pilot that impressed everyone and never went live
74%
Of AI value goes to 20% of companies
A small group captures nearly three-quarters of the measured economic value from AI. The leaders redesign workflows and put AI inside them, rather than adding more tools beside them.
Engineers who already know your constraints.
Production systems inside regulated, high-traffic and safety-critical environments — so the agent arrives already designed for the rules you operate under.
Finance
Document intake, onboarding checks, reconciliation and customer operations — where every automated decision has to survive audit, model governance and a regulator asking why.
Transportation
Operational data at network scale: disruption and asset workflows, customer correspondence and public-facing services whose demand arrives in surges rather than a steady line.
Entertainment
Media and content platforms: metadata enrichment, rights and archive workflows, and editorial support at a volume no team could read by hand.
Digital healthcare
Clinical and patient-facing systems where data protection, information governance and clinical safety decide what an agent may do on its own — and what it must always hand to a clinician.