AI should be assistance, and here is what that requires

AI is only safely useful in a services firm where five things are already true, and none of them are about the model.

A defined output. An approved source of truth. A template or playbook that says what good looks like. Human review before anything reaches a client. And a log of what was generated, from what, and who approved it.

Those five conditions are the whole of our position on AI. They are deliberately unexciting, and they are the reason we describe what we build as assistance rather than as anything more ambitious.

The mistaken diagnosis

The common reading is that AI in professional services is held back by model capability — that the drafts are not quite good enough yet, and that the next release will close the gap.

That has not been our experience. In the work we see, the constraint is not the quality of the generation. It is that the firm cannot say with confidence what the generation was based on.

A model asked to draft an engagement letter will produce something fluent from whatever it is given. If what it is given is a proposal document, a half-current rate card and a folder of similar letters, the output will be fluent and wrong in ways that are expensive and hard to spot — a rate that lapsed, a deliverable that was cut in negotiation, an assumption that belonged to a different client.

The failure is not hallucination in the usual sense. It is that the firm had no single approved record of what was agreed, so there was nothing correct to ground the draft in.

What each of the five conditions is actually doing

A defined output. Not “use AI in delivery” but “produce this document, of this type, for this reader, at this point in the process”. Undefined scope is how AI pilots end without a result anyone can name.

An approved source of truth. The specific records the draft may draw on, and the rule that it may draw on nothing else. If the engagement record carries the scope, the milestones, the rate card and the change history, the draft is grounded in the same data the invoice will later be built from. If it is grounded in a document library, it is grounded in history.

A template or playbook. The structure and the language the firm has already decided are correct. This is what stops the output being a new opinion each time, and it is also what makes review fast, because the reviewer is checking facts against a known shape rather than reading fresh prose.

Human review before the client sees it. Not advisory. The output is a draft, it is presented as a draft, and a named person accepts or edits it. Nothing generated goes to a client, or into a billing decision, without that step.

Audit logging. What was produced, from which records, by whom, and what was changed before it went out. This is the condition firms skip, and it is the one that decides whether you can answer a client or an auditor six months later.

Two mechanisms, described plainly

The first is drafting. An engagement letter assembled from the scoped milestones, deliverables and dates already recorded on the engagement — not from a blank page and not from last quarter’s letter. The saving is real but bounded: it removes the assembly, not the judgement. Somebody still reads it, and should.

Both run on Salesforce Agentforce, which is why the grounding question above is answerable at all: the assistant reads the same governed records the rest of the platform writes to, rather than a separate copy of them.

The second is internal analysis. A report on an engagement in flight, compiled from the time, cost, budget and milestone records already held against it: how effort is distributed, how budget is burning, where the timeline has moved, and which of those patterns has historically preceded trouble. It is written for the person running the business, not for the client, and it is deliberately unflattering when the data is unflattering.

What both share is that the grounding is the operating record, the output is a named artefact, and a person decides what happens next. Where that record lives is covered on our AI and analytics page.

Where the argument stops

Human review is what makes assistance safe, and it is also the hard ceiling on the time it saves. A firm buying AI in order to remove the review step is buying something other than what we have described, and should be told that plainly rather than discovering it in an escalation.

The bound also matters in the other direction. Plenty of firms get genuine value from AI on unstructured work — research, summarising a transcript, rewriting a section, thinking out loud — with no operating-record discipline whatsoever, and nothing here argues against that. The five conditions apply where a wrong answer is a commercial answer: pricing, resourcing, WIP and billing. That is a narrower claim than the market usually makes, and it is the one we are prepared to defend.

There is a third limit worth naming. Grounding a draft in the approved record only helps if the approved record is right. If your scope lives in a Word document and your rate exceptions live in an email thread, adding AI to that will produce confident output from an unreliable base — faster, and with more authority than it deserves. The work is upstream, in the scoping and quoting discipline that decides what the record says in the first place.

Assistance, and what that word is doing

We use “assistance” precisely. It means the output is named, the grounding is stated, and control stays with a person who can see both. It is not modesty and it is not hedging — it is the description that matches what the software actually does.

If your firm cannot currently say what a draft would be grounded in, that is the finding, and it is a more useful one than any pilot. Start with the record. The rest is downstream of it.