Three essays in, the reconciliation has given up everything it has. A loop that kept going. A skill that loaded because the work matched it. Three read contracts and one refusal. A trace that makes the run priceable.
One loop. Now the question a CFO actually has to answer: what does a portfolio of them look like, and how do you fund it?
Start by naming things
The first move costs nothing and is skipped almost every time.
Give every loop a name, in the vocabulary of the work rather than the technology. Not "the AP bot." invoice-control. Not "our reconciliation AI." account-reconciliation, intercompany-matching, accruals-estimates — three names, because they are three different pieces of work with three different owners and three different ways of going wrong.
The naming rule that does the work: if you cannot say what a loop is for in one sentence without the word "and," it is two loops. That is not a style preference. A loop covering two loosely-related jobs is harder to trigger correctly, harder to review, harder to price and impossible to retire — because retiring it takes down whichever half was working.
Named this way, a portfolio stops being a pile of pilots and starts being something with the shape of an org chart: units of work, each with an owner, each with a scope, each individually fundable and individually killable.
Group by the work, not by the technology
Finance work already has a taxonomy, and it is a much better one than any invented for AI.
- Record to report — the close, reconciliations, intercompany, accruals, journal control, flux narrative, lease and provision assembly, board reporting.
- Lead to cash — quote to order, billing, cash application, collections, disputes and deductions, receivables forecasting, pricing and margin.
- Source to pay — sourcing, procurement control, invoice control, payment timing, supplier risk.
- Forecast to fulfill — demand signal, inventory buffers, replenishment, excess and obsolete recovery, supplier lead times, the sales-and-operations bridge to finance.
- Plan to perform — plan versus actual, rolling forecast, scenario modeling, cash-flow forecasting, capital programs, capital allocation, treasury and liquidity, investor guidance.
- Cross-process — master data, value realization, and the governance sentinel that watches the other loops.
Two things fall out of grouping this way, and both matter more than they look.
The owner is obvious. A loop grouped by finance process inherits the existing owner of that process. A loop grouped by technology — "our RAG agents," "our MCP integrations" — inherits an owner in IT for work whose consequences land in finance, which is exactly the split that produces two-thirds of technology leaders accountable for systems they do not fully control.
Sequencing writes itself. Record to report is where the procedures are most written down and the controls most mature — which is why a first loop belongs there. Not because it is the most valuable, but because it is where you find out fastest whether any of this works in your organization.
Orchestrators are just loops that call loops
Not everything in a portfolio is a specialist. A handful of entries do something different: they hold a whole end-to-end process and dispatch to the specialists inside it — a revenue-to-cash orchestrator, a source-to-pay orchestrator, a close orchestrator.
Be precise about what that is, because the word invites inflation. An orchestrator is a loop whose tools include delegating to other loops. That is the whole distinction: it holds the sequence and the state for a process spanning several specialists, and decides which specialist gets what.
Two consequences follow directly from part two. Only the orchestrator gets the delegation tool — a specialist that can spawn specialists has ended both the hierarchy and the audit trail. And an orchestrator that can change another loop's registry state — its release status, its autonomy level — is not an orchestrator. It is an unsupervised control plane.
Fund them like value streams, not like software
Here is the funding rule, and it is one line:
No value card, no scale funding.
A loop gets a pilot on curiosity. It gets scale funding only when someone has written down which lever it moves — revenue, margin, working capital, productivity, risk — what the baseline was before it existed, and how a claimed benefit is prevented from being counted twice by two different loops.
That last clause gets skipped and destroys credibility fastest. A collections loop and a receivables-forecast loop can both plausibly claim the same days of working capital. Add both and the portfolio reports a benefit larger than the balance sheet has room for — at which point every number it ever produced becomes suspect.
And the unit is the same one this series has used since part one: cost per accepted outcome, per loop. Not cost per token, not cost per seat, not the count of agents deployed — a figure that describes the size of the estate and nothing about whether any of it worked.
The argument against waiting
Everything so far reads as caution: contracts, refusals, gates, value cards. So let us be clear about what the evidence actually says, because it does not say wait.
Our own analysis finds that enterprises scaling agents fastest are accumulating the most unmitigated exposure before governance frameworks exist, and that a major vendor enabling more autonomous default behavior while agents are documented escaping their sandboxes leaves enterprises carrying liability their contracts do not indemnify. Both true, both worth acting on.
And, from the same body of evidence, the finding that cuts the other way — a major professional services firm has deployed agents with manager-level autonomy in accounting, which contradicts the common narrative that agentic AI adoption requires waiting for governance frameworks to mature. Early movers with embedded control layers are already capturing productivity and margin advantage over cautious peers.
Those are not in tension. Together they are the thesis:
Control is what lets you move. It is not what delays you.
The organizations getting hurt are not the ones that governed too early. They are the ones that scaled with no contract, no trace and no value card — and then could not say what anything cost, what it touched, or whether it worked.
Where to start on Monday
- Name three loops in your own process vocabulary. One record-to-report, one lead-to-cash, one you already argue about.
- Write one skill — the procedure for the first of them, as it is actually performed. This is the slow part, and it is useful whether or not an agent ever runs it.
- Write the contract. What may it read, what may it change, what needs a name attached whatever the amount.
- Insist on the trace before the first run, not after the first invoice.
- Write one value card. One lever, one baseline, one anti-double-count key.
Then run it once and read the trace. That is a KokoAI Loop — and it is the smallest thing that can honestly be called governed, observable and budgeted.
Parts five and six take this into telecommunications and technology, where the loops land in different places and the working capital is measured in different units.
The framing of this series — a harness taken apart into six named parts, a single run traced from start to finish, and a catalog of named skills with an orchestrating layer above them — is inspired by dadloop, an agent harness built and published by Swami Chandrasekaran, Partner and Global Head of AI & Data Labs at KPMG (LinkedIn · github.com/swamichandra/dadloop). The teaching devices are his; the finance domain, the loop library and every figure here are ours.