Service 03

AI implementation

Retrieval over your documents, agents that complete real work, and automation wired into the systems you already run, built to be handed over, not rented back to you.

What we build

Retrieval over your own material

Question answering, search and drafting grounded in your documents, policies, contracts, case files or records. The version that survives a security review: permissions resolved per user at query time, every answer cited to a source someone can open, and a corpus with an owner and a lifecycle so superseded documents stop generating confident wrong answers.

Document and data extraction

Turning unstructured material into structured fields you can act on: invoices, forms, correspondence, reports, scanned records. Usually the highest-return AI work available to an organisation, because the current process is a person retyping things, and because accuracy is measurable rather than a matter of opinion.

Workflow automation

Classification, routing, triage, summarisation and drafting, wired into the systems where the work actually happens rather than into a separate tool people have to remember to open. An AI feature nobody opens is worth nothing.

Agents, where they genuinely fit

Multi-step systems that plan and use tools, for tasks that vary meaningfully each time. We build these with tight tool permissions, human approval at the points where an action is consequential, and full logging of what was decided and why. We will also tell you when a deterministic pipeline would do the job better, which is more often than the market implies.

How we build

Real data early

We work against a slice of your real material under a proper agreement, from as early as your governance allows. Pilots on synthetic data prove that a demo can be built. They tell you almost nothing about whether the system will work, because real corpora are messier, more inconsistent and more sensitive than the sample anyone prepares for a demo.

Evaluation as infrastructure

Before we tune anything, there is a test set of real questions with reviewed correct answers. Every change is measured against it. This is what lets you answer "how accurate is it?" with a number when the person who has to approve the rollout asks, and they will.

Boundary first

The data classification decides the architecture, and we settle it in week one. Retrofitting a boundary into a system built without one is not a modification; it is a rebuild, and it is the most expensive mistake in this field.

Built to be handed over

Tests, CI, infrastructure as code, architecture decision records, a runbook, and sessions with your engineers. We are trying to make ourselves unnecessary. Clients tell us afterwards that this is the part that mattered most, and it is the part most suppliers quietly skip.

How an engagement runs

Weeks 1-2: Scope and boundary. Confirm the use case, settle the data classification and deployment tier, build the evaluation set, agree what success means and who signs it off.

Weeks 3-8: Build. Working software from the first fortnight, reviewed with you every two weeks against the evaluation set. Integration work runs in parallel, because it is usually the long pole.

Weeks 9-12: Harden and hand over. Security review, monitoring, load and failure testing, documentation, handover sessions, and a defined support period while your team takes ownership.

What we will not do

Build something we think should not exist. If the assessment says a rule engine or a fixed query does the job, that is what we will recommend, and it will cost you less.
Put our own licensed middleware permanently between you and your systems.
Ship without an evaluation set, because then nobody, including us, knows whether it works.
Claim a human-in-the-loop control where the human would be approving faster than they could possibly read.

Questions we get asked

Do you build agents?

Where they earn it. An agent (a system that plans, calls tools and acts across several steps) is the right shape when the task genuinely varies each time. When the task is the same every time, an agent is an expensive and unpredictable way to run a workflow that a deterministic pipeline would handle more reliably and far more cheaply. We are happy to build either; we will tell you which one your problem is.

How do you handle permissions in a document assistant?

Access control from the source systems is carried into the index as metadata and enforced before retrieval, against the identity of the user asking, not a service account. This is the single most common reason internal assistants fail their security review, and it is close to a rewrite if retrofitted. We design it in from the first week.

What about accuracy? Our people cannot check every output.

Then the system has to be honest about uncertainty and cite its sources, so checking is cheap when it matters. We build an evaluation set from your real material, measure against it, and include questions the system should refuse to answer. "I could not find that in the documents" is a first-class outcome, not a failure, and refusal behaviour is the first thing to degrade when a system is tuned for helpfulness.

Will this integrate with SharePoint, Teams, our CRM or our case management system?

Generally yes. Most of the work in a real implementation is integration and data preparation rather than model work: extraction from awkward formats, permission mapping, and handling the documents that are scanned images rather than text. We scope integration explicitly rather than treating it as an afterthought, because that is where the time actually goes.

Do you hand over the code?

Yes, entirely. It goes in your repository under your licence, with no proprietary runtime of ours sitting in the middle and no per-seat fee on work you paid us to build. If we part company, everything keeps working.

How do you price a build when scope is uncertain?

Fixed fee where scope can be fixed, which is most of the time after an assessment. Where it genuinely cannot, we propose a short paid discovery to get to a scope we can quote, rather than quoting a number we would have to revise upwards later.

Start with a straight answer

A 30-minute call, no pitch deck. Tell us what you are trying to do and we will tell you whether AI is the right tool, what it would take, and what it would cost, or that you should not bother.

Book an intro call See how we price work