Finance — Charter

Finance keeps the company solvent and tells everyone else the truth about the numbers. In an AI-operated company it gains a cost line that behaves unlike any other: inference spend scales with usage, not with capacity, so it grows quietly and nobody owns it by default.

Automation in finance also carries a different risk profile from everywhere else. Elsewhere a mistake is embarrassing. Here it is a misstatement, a mispayment, or a control failure.

What this role owns#

Planning and forecasting. What we expect to earn and spend, and how confident we are.

Cash. Runway is the only number that ends the company if it reaches zero.

Unit economics. What it costs to serve one customer, deliver one project, run one AI feature. Without this, "AI is expensive" and "AI is cheap" are both unfalsifiable.

Cost attribution, including AI spend broken down by system and by outcome.

Controls. Approval thresholds, segregation of duties, and the fact that these apply to agents exactly as they apply to people.

KPIs#

AI spend, specifically#

The cost line most organisations mismanage because it is new:

• Attribute it to a system and an outcome, not to a single "AI" budget line. "We spent X on AI" is not a manageable number.
• Measure cost per completed task. An agent that wanders is expensive and unreliable, and the cost metric detects it before the quality metric does.
• Set hard caps per task, treated as controls rather than warnings.
• Watch context growth. Cost scales with how much context is sent, not just how many requests are made — and prompts grow by accretion unless someone measures them.
• Separate experimentation from production so a research spike does not look like a cost trend.

AI agents in this function#

Transaction categorisation agent — codes routine transactions; exceptions to a person.

Reconciliation agent — matches records between systems and surfaces only the breaks.

Forecast assistant — builds a base forecast from history and flags where the assumptions are doing the heavy lifting.

Anomaly detection — spend outside pattern, duplicate invoices, unusual vendor activity.

Cost attribution agent — assigns infrastructure and AI spend to systems and customers.

What stays human — and this list is firmer than in other departments: approving payments, signing off statements, accepting an audit position, changing bank details for a supplier, and anything that moves money. A supplier emailing new bank details is the most reliably profitable attack there is; verification happens out of band, by a person, on a channel the requester did not choose.

SOPs#

• Payment approval — thresholds, segregation of duties, and out-of-band verification for any change to payment details.
• Month-end close — sequence, owners, and what blocks completion.
• Reconciliation — daily where possible. Breaks found daily are investigable; found monthly they are archaeology.
• Cost review — monthly, top ten lines, AI spend separated, with a named owner per line.
• Forecast revision — when and on what evidence, so revisions are decisions rather than drift.

Templates#

Budget and forecast · cost attribution model · Project Plan budget section · capital request · payment approval record.

Depth on the constraints in FinTech — particularly why money must never be a floating-point number.

Workflows#

In: invoices · sales forecasts · payroll · infrastructure and AI cost data · capital requests.

Out: management accounts · forecast · runway position · approved or refused spend · unit economics by product.

Handoffs: CEO for capital allocation · COO for capacity funding · Sales for forecast inputs · DevOps and CTO for infrastructure and AI cost.

FAQ#

Can AI do the bookkeeping?#

It can categorise, reconcile and flag anomalies well, which removes most of the mechanical work. Approval, judgement and sign-off stay human — not because a model could not classify correctly, but because someone must be accountable for the statement.

How do we budget for AI when usage is unpredictable?#

Budget per outcome rather than per period, and enforce caps. "We will spend up to X per completed task, and no more than Y per month" is enforceable in a way that a single annual number is not.

What is the most common financial control failure with AI?#

Nobody owning the spend. It accumulates from thousands of small decisions made by people with no visibility of what theirs cost. Assigning each cost line to a person who reviews it monthly finds more waste than any tooling will.

Should agents have access to financial systems?#

Read access for reconciliation and analysis, yes, and it is genuinely valuable. Write access to anything that moves money should be no — and if an exception is made, a human approval gate goes in front of it, without exception.

What else is coming for Finance