Engineers who believed AI deserved better delivery

Every engagement begins with a fixed-fee discovery, because a fixed price quoted before anyone has seen your data is either padded or fiction. Discovery produces a committed scope, timeline, price and expected ROI — and occasionally a recommendation not to proceed, which we have delivered often enough that clients now cite it as the reason they trusted the rest.

Delivery runs in weekly cycles with a live demo on your real data. Not a status deck, not a percentage-complete figure: working software, running against the actual inputs it will face in production. This is the single practice that most reduces the risk of a late unpleasant surprise, because it makes the gap between "works in the demo" and "works on your data" visible in week two instead of week twelve.

Teams are three to five senior engineers with a named lead who is accountable to you and present in every demo. We do not staff a junior bench behind a senior sales team, and we do not grow teams to increase billing — coordination cost rises faster than throughput, and AI delivery rewards depth over headcount more than most software work.

Every project ends with a retrospective that changes how we work. The evaluation harness pattern, the shadow-mode rollout, the pre-pilot checklist and the parallel-run methodology all came out of engagements where something went wrong and we wrote down why.

A consultancy is defined as much by its declines as its deliveries, so it is worth being specific.

We do not build systems designed to make people believe they are talking to a human when they are not. Our voice agents disclose themselves on every call, in every jurisdiction, regardless of whether local regulation currently requires it.

We do not build cold-calling or collections voice agents, and we do not build employee surveillance systems. The technology is capable of both; the regulatory and reputational exposure is not worth it for our clients or for us, and we say so before the conversation gets far.

We do not build systems that make consequential decisions about individuals — hiring, credit, insurance, clinical, disciplinary — without documented human oversight. We will build the system that prepares the decision, evidences it, and makes the human faster. We will not build the one that removes them.

And we decline work we do not believe will reach production. That usually means the data is not there, the process changes too often to encode, or no owner exists on the client side. Taking that work would be profitable in the short term and is the fastest available way to acquire a portfolio of dead pilots.

Ninety-seven percent of our clients continue past their first engagement, and the reason is not loyalty — it is that the first system worked and there is a second workflow behind it.

The mechanism is deliberate. Every system launches with agreed KPIs and a dashboard tracking them, so the question of whether it worked has an answer rather than an opinion. When a client is deciding whether to fund a second project, they are looking at a number from the first one. That is a much better position for both of us than a relationship built on rapport.

The second reason is that we build for handover. Your team owns the code, sits in on delivery, and gets runbooks and working sessions on the parts they will operate. Clients who could run everything themselves after the first project frequently choose to work with us on the next one anyway, and that is the only version of retention worth having.

The third is scope discipline. We would rather deliver one workflow that reaches production than three that reach a demo, and we say so during scoping even when the larger number would be more profitable in the quarter. Clients notice that the second engagement is proposed on the evidence of the first rather than on momentum.

Every project ends with a retrospective that is allowed to change how the company works, and four of our standard practices exist because something went badly enough to warrant one.

Evaluation-first delivery came from a document automation project in 2021 that spent eleven weeks in a loop: someone would find a bad extraction, confidence would wobble, another round of tuning would follow, and someone would find another bad extraction. The system was never bad enough to kill and never trusted enough to ship. We now refuse to start a build without a golden dataset and a written threshold agreed by the business.

Shadow-mode rollout came from an agent that was granted write access to a scheduling system one category too early. Nothing catastrophic happened — the guardrails held — but the disagreement rate with human operators was higher than anyone had measured, because nobody had measured it. Every agent and voice system we deploy now runs proposing-not-acting against live data until the disagreement rate is both low and understood.

The parallel run came from a client asking a question we could not answer: how good is the process we are replacing? We now measure the manual baseline error rate before cutover in every automation engagement. It has never been zero, it is typically between one and four percent, and it reframes the decision from "is the AI perfect?" to "is it better, and are its errors more detectable?"

The pre-pilot checklist came from counting. Of the engagements that stalled, almost all were missing a named production owner, a written quality threshold, or a unit-economics model. Those three questions now get asked in the first discovery call, and we have walked away from work where the answers did not exist and could not be created.

None of these are proprietary. We publish all four, in detail, on this site and in our writing — partly because the industry is better when they are standard practice, and partly because a client who applies them and then hires someone else has still had a better outcome than one who did not.

Three things are changing the shape of enterprise AI delivery, and we are building for them deliberately rather than reacting.

Agents are moving from one bounded workflow per engagement to portfolios that share a runtime, a tool registry, an evaluation harness and a governance layer. The engineering problem shifts from building an agent to operating many, which is why our recent platform work has concentrated on orchestration, observability and cost attribution rather than on agent logic.

Regulation is arriving on a timescale that matters. The EU AI Act’s obligations are phasing in, India’s data protection regime is being operationalised, and sector regulators are moving faster than the general frameworks. We have invested in making compliance evidence a by-product of the architecture, because the alternative — assembling it by hand before an audit — does not scale past the second system.

And model economics keep improving in a way that changes which projects are viable. Workflows we declined on cost grounds two years ago are now straightforward. We re-score client portfolios quarterly for exactly this reason: an initiative that was infeasible in January is routinely obvious by September, and a strategy that is not revisited is a document rather than a plan.