AI readiness does not start with AI

The precondition for useful AI in a services firm is not a model or a licence. It is whether the firm can trust its own operating record without rebuilding it every month.

This is not a caution about AI. It is a statement about where the constraint actually sits, and it is testable in about twenty minutes without buying anything.

The mistaken diagnosis

The common reading is that AI adoption is a tooling decision. Choose an assistant, roll it out, train people, measure the time saved.

What that misses is what the assistant is reading. If delivery, time, cost and billing data lives across five systems, is entered late and is reconciled monthly, an assistant sitting on top of it can summarise, draft and search — but it cannot tell you whether a project is recoverable, because nothing in the record establishes that.

The same fragmentation shows up in the platform research. MuleSoft’s 2025 Connectivity Benchmark Report, summarised by Salesforce, put the share of connected applications at 29% across 1,050 surveyed enterprise IT leaders, and named that disconnection as a direct constraint on the accuracy and usefulness of AI agents. Those are large-enterprise figures, so read them as directional rather than as a benchmark for a hundred-person firm. The mechanism is the same at any size.

Six things to check before anything else

A useful readiness check has nothing to do with models. It has six dimensions, and each one is a question about your own operating record.

Commercial continuity. Does what was sold arrive in delivery as structured, usable inputs — scope, assumptions, rates, approvals — or as a PDF and a meeting.

Workflow discipline. Are approvals states on a record with an owner and a timestamp, or are they email threads that someone can locate if pressed.

Data trust. If two people ask the same question of the system this afternoon, do they get the same answer, and would either of them act on it without checking.

Commercial control. Do billing rules exist as data that the system applies, or as clauses that experienced people interpret consistently because they have been here a while.

Management discipline. Can WIP be interrogated on any given Tuesday, or is it a month-end construction. This is the one that most reliably predicts the others, and the argument behind it is set out in WIP: register or argument.

Automation readiness. Is there a step in your chain that runs the same way every time, on structured inputs, with a defined output. If not, there is nothing for automation to attach to.

Score these with your CFO, your delivery lead and whoever owns your CRM, separately. The disagreement between the three scores is more informative than any of them individually — it is the gap between the leadership narrative and the operating reality, measured.

What the record has to carry

The mechanism behind all six is the same and it is unglamorous.

The engagement holds its commercial rules from the point it is quoted, so time entered against it inherits them rather than being classified later. Approval is a state, not a message. Change orders attach to the engagement they change and are priced before delivery. Each WIP entry links back to the work, the approval and the contract clause that permits billing it.

Once that holds, an analysis has something to stand on. DAY ONE’s reporting and AI features read from that governed record rather than from a copy: the output is named, its grounding is the source records it links back to, and a person reviews and approves it before it reaches a client or a ledger. The specifics are on our analytics and AI page, and the access, permission and audit model underneath is on the security page.

Where the claim stops

This argument is bounded and the boundary matters.

A firm can get real, immediate value from AI on unstructured work with no operating-record discipline at all. Drafting, summarising, first-pass research, meeting notes, code, document review — that value is available today, it does not depend on your WIP being traceable, and telling people otherwise to sell a governance project would be dishonest. If that is what you want from AI, go and get it.

The claim applies to AI that touches pricing, resourcing, WIP or billing. In that territory a wrong answer is a commercial answer — it goes to a client, or into a forecast, or onto an invoice — and an assistant working from a record that has to be rebuilt monthly will produce those answers with exactly the same fluency as correct ones. That is the specific risk, and it is a data problem rather than a model problem.

There is a second cost worth naming. Getting the record into that state is months of unglamorous structural work with no demo at the end of it, and it competes for attention with AI projects that look far more impressive in a board pack. Firms routinely choose the demo. That choice is understandable and it is why the constraint persists.

Where to start

Run the six questions this week, with three people, separately. Fix the dimension where the three of you disagree most, because that is where your operating reality and your reported reality have come apart. Then reconsider the AI question — it will be a smaller and much clearer decision than it is today.