Groundwork: The firms preparing for agentic AI right now are doing three things that will put them years ahead

While most of the market is still debating chatbots, a handful of firms are laying the groundwork for agentic AI: systems that don't just answer questions or act as an assistant, but carry out whole pieces of work from beginning to end. And there is the phrase we hear in every one of those boardrooms: "don't worry, there's a human in the loop". It's the sentence that soothes many AI conversations I sit in.

So the question I've started asking back is along the lines of: how can we put a human in the loop if we're not sure what the loop is? In most organisations, the workflow the agent is meant to run has never been written down. It lives as instinct in the heads of three senior people.

Prepared firms are doing groundwork now that will give them an enviable lead on the pack. Let's break it down a bit so we can see what "the agentic revolution" means for you, me, and our teams, next Tuesday afternoon when real work is due.

First, what this looks like when it works

Forget vendor announcements about "running thousands of agents"; that tells us nothing about what they're using those agents for or the quality of the output. What agentic work looks like close up, in workflows I see being built right now, is this:

Communication that tailors itself to every reader. Picture a firm that holds regular investor or client calls. After each one, an agent reads the transcript and logs what each person actually asked about: one investor pressed on cash runway, another on churn, a third only ever cares about the pipeline. When the next update goes out, each version leads with what that reader cares about and answers the question they asked last quarter before they ask it again. Every round of feedback is fed back in, so the reports get more precisely tuned with each cycle. One team, communication that feels handwritten to every recipient, at scale. That used to be impossible at any headcount.

A month of authentic content from work already happening. Your workshops, client sessions and talks are already being transcribed. An agent combs those transcripts for the strongest moments, strips out anything confidential, reshapes each one in your house voice, and queues it for a human's final polish before it's posted. Nothing reads as machine-made, because the raw material is real people saying real things; the agent just does the mining, cleaning and shaping that nobody ever had time for.

Proposals drafted from the conversation itself. A sales call ends; the transcript goes to an agent that returns a fully formed proposal in house style, built around what the client actually said, with the right case study already chosen. The person who used to lose an evening to it now spends fifteen minutes sharpening judgement into it. In a corporate finance setting, the same pattern pre-populates a third to half of a new business proposal from public filings, prior correspondence and the firm's own precedents, leaving the analytical judgement, and only the judgement, to the humans.

The pipeline that updates itself. One team stopped typing notes into a CRM nobody read. Instead, agents analyse transcripts from the real client conversations, and maintain the relationship map themselves: who's warm, who's gone quiet, where the risk is building. It proactively surfaces the highest priorities without anyone having to ask. The client picture is drawn from what actually happened, not what someone remembered to log on a Friday.

And at the giant end of the scale, the same logic holds: Walmart hands a slice of supplier negotiations to an agent that closes around two-thirds of the deals it attempts, and when researchers asked, most suppliers said they preferred it, because it responds in hours and never gets tetchy.

Notice the pattern across all five.

The work gets faster and better and more personal, all at once.

And behind every one sits the same unfashionable preparation, which is the subject of this newsletter, because it's the part you can start this week, that has real, tangible, measurable value in days, weeks, months - not at a nebulous point in the agentic revolution future.

Thing one: they are mapping their workflows

A workflow is simply what happens, in order, to get a piece of work done. Who receives the request, what they check first, what they produce, who reviews it, what happens when something looks odd, and the quality control that happens all the way along.

The catch: most of us would likely struggle to write our own workflows down, because we run them on autopilot. Daniel Kahneman described two modes of thought: the fast, automatic kind that handles the familiar, and the slow, deliberate kind we save for the new. Years of experience push our best work into the fast lane.

It's the same reason you can drive home and remember nothing of the drive; the route has sunk beneath conscious thought.

Your best underwriter, editor or engineer performs a dozen judgements before lunch that they could not list if you asked.

That works for humans, because humans learn by watching; sit a trainee beside your best person for six months and the invisible steps transfer by osmosis. Agents don't learn by watching. They learn by being told, and they cannot infer the step you left out. So the first act of groundwork is making the invisible visible: sitting with the people who do the work and writing down what actually happens, in order, including the two-second glance, the unwritten rule, the exception everyone knows and nobody documented. This, incidentally, is how you draw the loop. The firms doing it are writing the recipe book of their own expertise, and it turns out to be valuable before any AI touches it.

Thing two: they are deciding, stage by stage, where AI adds what

With the workflow on paper, a better question becomes possible. Not "should we use AI here?", which is too blunt to be useful, but: at this stage, what would AI actually do? Speed it up? Deepen it? Raise its quality? Free the human for something better? The answer differs at every stage. And this is where leadership matters deeply, because what you choose to optimise for shapes everything: a firm optimising for cost finds one set of uses; a firm optimising for quality, or for taking the drudgery out of its people's week, finds quite another. The workflow map is the table on which that choice gets made, stage by stage, rather than in the abstract.

Thing three: they are automating rung by rung, not in one leap

This is the part I most want to share, because it's the actual mechanics of getting to agentic AI safely when we're talking regular workflows across business.

They don't switch a workflow over to an agent, beginning to end, and hope it performs as highly as the human who used to do it.

Instead, they start on the first rung: AI runs stages one and two, and a human reviews everything, edits it, and, crucially, puts the corrected version back into the system so it learns what good looks like here. Think of a driving instructor's car with dual controls. The learner drives; the instructor's foot hovers; every correction is a lesson. Over weeks the corrections dwindle. When the outputs are reliably right, and you know because you've been measuring, the human steps back from that stage, the agent runs stages one, two and three, and the human review moves one rung up the ladder. Repeat along the chain.

Done properly, this is deliberate work, and it takes time. Verification checkpoints designed into every stage. Humans in the loop who actually know what good looks like at their rung, which is a skill that has to be built. Transparency, so the whole team can see what the agent did, what was corrected and why, and learn from it. That scaffolding is precisely what the failed projects skip: Gartner expects over 40% of agentic programmes to be scrapped by 2027, and MIT found 95% of AI pilots return nothing measurable. Same technology as the firms above; missing groundwork.

Now let's watch it happen to one real piece of work

Rather than describe this in the abstract, let me give you an example in a sample week. The scenario is a research briefing for a big decision, in this case whether to enter a new market, but every organisation has its own version of this week: an acquisition target, a major pitch, a product launch. Swap in yours as we go.

Monday, 4pm. The brief. The old version was an email: "Can you look into the German market? Board wants a view Thursday." Everyone downstream interpreted it differently. In the new version, the MD spends fifteen minutes answering the agent's intake questions: what decision are we making, what criteria matter, what would change our mind, who reads this. Fifteen minutes of precision at the top, because the agent needs it, and every stage below inherits it.

Monday, 4.30pm. Scoping. The agent comes back with 212 companies that could serve as comparators, grouped by how closely they resemble us. The old constraint, three comparators because a human had three days, has vanished. The head of research spends twenty minutes cutting 212 to 58, and that cut is pure judgement: which of these are actually like us, which markets rhyme with ours. The machine proposed; the human pruned.

Monday night. Gathering. While everyone sleeps, agents pull filings, pricing pages, customer reviews, hiring patterns and local coverage across the 58, in four languages. Nobody's evening was spent in browser tabs. The dead hours became working hours.

Tuesday, 8am. Screening. Waiting in the inbox: all 58 scored against Monday's criteria, with 14 flagged for human attention and a sentence explaining each flag. In the old workflow, whatever sat in the unread pile stayed unread, and everyone hoped it didn't matter. Now nothing important dies unread; human eyes go exactly where they're needed; where they're most valuable.

Tuesday morning. The deep read. The analyst spends the whole morning reading those 14 properly. Not skimming at 11pm against a deadline; reading, with time to think. Notice where the freed time went: to the most human part of the entire job.

Tuesday, 2pm. Synthesis. The agent produces a first draft in the house template, every section where it should be. The analyst's afternoon goes into reshaping the argument, not wrestling the formatting. The energy moves from how it's laid out to what it means.

Tuesday, 4pm. The challenge round. The analyst turns the agent on its own work: give me the strongest case against this recommendation. What would our most sceptical board member ask? What's missing? Two weaknesses surface; the draft gets stronger. This stage did not exist in the old workflow, because a tired analyst didn't have time to argue with their own report. It's not a faster version of an old thing. It's a new thing.

Wednesday morning. Verification. Every claim in the draft is accompanied by its credible, hyperlinked source ready for cross checking. The analyst checks the six claims the recommendation actually leans on; one doesn't survive contact with its source and comes out. Two global firms refunded or withdrew reports this year over invented citations. This half-day is why prepared firms won't be joining them.

Wednesday afternoon. The briefing. What goes to the board is two pages, shaped around the questions this particular board asks, with the full depth one click below for anyone who wants it. The old forty-page pack, of which six pages got read, is gone. The scarcest resource in the room, attention, finally gets respected.

Thursday. The decision. The call is made by humans, exactly as it always was. But it's made standing on evidence from 212 companies instead of three, it took three days instead of three weeks, and the reasoning is captured, so that next year, when someone asks why we decided this, there's a full documented answer.

Read back through that week and notice what the machine took: the gathering, the screening, the formatting, the first drafts, the source-chasing. And notice what the humans kept: the brief, the pruning, the deep read, the argument, the call. Nobody was replaced. The work was re-divided, the quality rose at nearly every stage, and one stage that never existed before appeared.

A head of research once told me he models three comparable companies per deal, because three is optimal: enough to compare, not so slow it stalls the deal. Perfectly rational, for a world where a human did the searching. Then overnight agents made it three hundred comparisons. And so then, something more interesting than speed happens: the definitions themselves started to shift: what counts as a good comparison changed. What counts as an informed decision changed. What counts as an engaging report changed, because a report can now anticipate the questions the room will ask. AI doesn't just do the old work faster. It makes new work possible, and the standards reset around whoever gets there first.

Now widen the lens

That was one workflow, for one person. Every individual in your organisation runs a dozen. Map them, enrich them stage by stage, climb the ladder rung by rung, and even a one percent improvement per stage compounds the way interest does: invisibly for a while, then unmistakably. Across a team, that's a different quarter. Across an organisation, a different cost base and a different quality of decision. Across a sector, it's the gap between the firms setting the standards and the firms wondering why clients' expectations keep rising.

The organisations doing this groundwork now aren't announcing it loudly. It will be met, and seen, in twelve to eighteen months, in their proposals, their prices, their strangely personal communications and their speed, and it will look like magic. It's a recipe book, a ladder, and patience.

If you'd like somewhere to start, try the exercise our clients do in their first workshop: pick one workflow that matters, walk it through as ten real stages with the people who actually do the work, and mark at each stage where a machine could gather, screen, draft or check, and where the human judgement must stay. That single afternoon is worth more than any extra licence to buy this year.

I read every reply, so tell me: which workflow would you map first? And if someone forwarded this to you, you're very welcome to subscribe.