Work with me

Treasury and finance workflows, redesigned — with AI where it holds up

I take one slow, manual, exception-heavy workflow — bank statement processing and reconciliation, the daily cash position, payment approvals in SAP BCM, hedge accounting, month-end — and redesign it end to end: what stays a rule, where AI reads or drafts, and which decisions stay with a named person. The same method works elsewhere in the enterprise; this is where I move fastest, because SAP, finance and treasury programmes are the ground I know best.

This page is for teams with a finance or treasury workflow to fix. Here for the daily steps instead? They go out on LinkedIn and X; the About page says what they are.

How it starts

  1. A free 30-minute call. Send a short request and I'll reply within two working days to arrange it. Tell me what the workflow is and what makes it painful, and I'll say honestly whether I can help — NDA first if you need one.
  2. A written scope within a week. Which engagement fits, what is in and out, and a fixed fee.
  3. The work. Weeks, not months. Remote by default, on site where the work needs it.
Request a free 30-minute call

The main engagement

AI Workflow Redesign Sprint

One complex workflow. Redesigned end to end around AI.

For a workflow you already know is a problem: slow, manual, exception-ridden, or quietly expensive.

We map how the work actually runs — not the process diagram, the real thing, including the parts nobody documents — and find where the elapsed time and the rework genuinely sit. Then we separate what must stay deterministic (calculations, limits, postings, segregation of duties) from what AI does well (reading, synthesising, classifying, drafting, spotting the exception) from the decisions that need a named human. What comes out is the version of the workflow that would be designed today: one that owns its own state, handles the exception path instead of ignoring it, and can be audited.

Before

  • Work arrives by email and spreadsheet, and someone chases it
  • Most of the elapsed time is waiting, not working
  • Exceptions are handled by whoever remembers how
  • The reasoning behind a decision lives in someone's inbox

After

  • Input is read and structured on arrival; the chase is the system's job
  • The deterministic core is still deterministic — and still auditable
  • Exceptions are a queue with owners, thresholds and escalation
  • Every decision carries its evidence, its author and its version

What it covers

  • Current state, with the time and rework where they really are
  • The deterministic core that must never be handed to a model
  • Where AI reads, drafts, classifies or proposes — and where it stops
  • Decision rights: who approves what, on what evidence
  • The state the workflow has to hold, and the exception path
  • The architecture and controls the redesign implies
  • A business case in cycle time, rework and risk, using your numbers
  • A roadmap: what to build first, and what to leave alone

You end up with: A redesign your team can build against, with the trade-offs written down rather than discovered in production.

See what a workflow redesign delivers

An illustrative excerpt from a workflow redesign specification, using a fictional bank reconciliation process.

Read the sample specification

Illustrative only. The scenario, identifiers and amounts are invented. This was not produced for a client and reports no results. A real specification uses your rules, systems, limits and controls.

Bank statement reconciliation: handling unmatched transactions

A. Current workflow

  1. Bank statement received
  2. Automated matching
  3. Unmatched item
  4. Manual investigation
  5. Resolution approval
  6. Clearing or follow-up

The problem: Investigation depends on fragmented context and undocumented decisions. To explain one unmatched line, an analyst checks open items in the accounting system, a shared spreadsheet of known exceptions and old email threads. The reasoning behind the resolution stays with that analyst, so the next similar item starts from nothing, and an auditor can see the posting but not why it was made.

B. Proposed workflow

  1. Deterministic matching rules. Configured rules compare amount, currency, value date and reference with open items, including permitted combinations of invoices and credit notes. A transaction that satisfies an approved rule posts straight through, with no model and no manual approval. Rules change only through change control, and each result records the rule version.
  2. Evidence gathering and context assembly. For each unmatched line, a script collects the same pack: the statement line, candidate open items, the customer's recent payments, related remittance messages and earlier resolutions of similar items. Read-only. A source that is unavailable is listed as missing, not worked around.
  3. AI-assisted investigation. A model reads the unstructured parts of the pack, such as remittance text, then summarizes them and may suggest an explanation, citing the passage each point rests on. It sees the pack and nothing else. It cannot post, clear, change a match, close an item or decide an accounting treatment, and a suggestion without a citation is not shown.
  4. Human decisions and approvals. For an unmatched transaction, a named analyst accepts, amends or rejects the proposed resolution, and states the reason. Exception postings and financial adjustments go to an authorized approver within their limit, who cannot be the person proposing them.
  5. Auditable resolution. Standard transactions post under their approved rule. Exception postings and adjustments post once the applicable approval is recorded. Each item keeps its evidence, decisions and posting together. Proposed audit design: an append-only history of decisions and events, with evidence and rule versions preserved. A correction is recorded as a new action linked to the one it corrects, so no earlier decision is silently overwritten.

AI never decides that a transaction is safe to clear. Only a configured rule or an authorized person can.

C. Decision ownership

SituationWho decidesControl boundary
Standard transaction satisfying approved matching rulesDeterministic rulePosts straight through, without AI or manual approval. Changing a rule needs change control.
Multiple possible matchesAnalystThe system lists the candidates. Nothing clears until a person selects one.
Missing reference informationException queue, with a named ownerThe item waits with an escalation date. No match is inferred from the amount alone.
Unstructured remittance explanationAnalyst, assisted by AIAI may summarize and suggest, with evidence. The suggestion is not a decision.
Final exception posting or financial adjustmentAuthorized humanWithin their limit, and not the proposer. Neither AI nor automation can approve.
AI unavailable, or its output fails checksAnalystThe item continues from the evidence pack alone. Nothing waits on the model.

D. Example exception

Statement STMT-2041, line 17: an incoming credit from customer C-1042.

  • Fact given in this fictional scenario
  • Assumption would have to be confirmed against the real process
  • Proposed the design being recommended
  1. What triggered the exception
    • Fact The line is for EUR 48,200.00 and its reference reads “INV 90317 90318 LESS CREDIT”. The configured matching rules, including permitted combinations of invoices and credit notes, find no valid match for this amount and these references, so the line goes to investigation.
  2. What evidence is available
    • Fact Open items for C-1042: INV-90317 for EUR 31,000.00, INV-90318 for EUR 18,500.00, and credit note CN-5521 for EUR 1,000.00.
    • Fact An email from the customer's accounts payable team: “Paying both less CN-5521. 300 withheld pending the damaged pallet claim.”
    • Fact No claim for damaged goods is on file for this customer.
  3. What the system can determine
    • Proposed By arithmetic alone: EUR 49,500.00 invoiced, less EUR 1,000.00 credit, is EUR 48,500.00 expected. The EUR 48,200.00 received leaves EUR 300.00 short.
    • Assumption The configured payment tolerance is EUR 50.00. The shortfall is above it, so no rule may clear this line.
  4. Where AI could help
    • Proposed Summarize the email, quote the sentence giving the reason, and suggest allocating the payment to both invoices and the credit note, leaving EUR 300.00 open.
    • Proposed Point out that the email mentions a claim and the evidence pack contains none. The claim stays unverified, and the suggestion says so. It does not propose an accounting treatment.
  5. What remains for the human decision-maker
    • Proposed The analyst decides whether the message is genuine and the allocation is right.
    • Proposed The EUR 300.00 stays an identifiable open receivable. An authorized analyst classifies it and routes it for dispute resolution or collection. The unverified claim is not treated as an approved deduction.
    • Proposed Any adjustment or write-off needs the applicable authorization.
  6. What gets recorded for audit
    • Proposed The statement line, the evidence pack as it stood, and the rule result with its version.
    • Proposed The summary and suggestion shown, with the model and prompt version, and whether they were accepted.
    • Proposed The decision and its reason, who made and approved it and when, the posting reference, and how the residual was classified and routed.

E. First validation experiment

Before anything is automated, the assistance runs in shadow mode: alongside the current process, with no access to post or clear anything.

Hypothesis
  • With the evidence pack and AI summary in front of them, analysts reach the decision they would have reached anyway, with less searching, and the suggestions do not lead them to a wrong one.
What to test
  • Build packs and suggestions for a fixed sample of real unmatched items over an agreed period.
  • Analysts resolve every item through the current process. For part of the sample, they record their decision before seeing the suggestion.
  • A reviewer compares the two afterwards.
Evidence to collect
  • For each item: the analyst's decision and reason, the suggestion, and whether each citation supports what it is cited for.
  • Sources that were missing from the pack.
  • Time spent investigating, with and without the pack, against a baseline measured first.
  • Why the analyst and the suggestion disagreed, wherever they did.
Success criteria
  • Agreed with the process owner before the first item is run, using the measured baseline. They are left blank here on purpose: a number without that baseline would mean nothing.
  • They cover at least: citation accuracy, how often a suggestion agrees with the final decision, whether wrong suggestions are caught, and investigation time.
Stop conditions
  • A suggestion states a fact or citation that the pack does not contain.
  • Any route is found by which the assistance could write to the accounting system.
  • Data reaches anyone or anything outside the agreed access.
  • Analysts begin accepting suggestions without checking the evidence.
Talk about your workflow →

If a sprint isn't where you are yet

Not sure which workflow?

AI Workflow Opportunity Scan

We go through the real workflows in a function and assess each against the things that decide whether AI changes its economics: how often it runs, where the human hours go, how much of the input is unstructured, how repeatable the judgement is, the exception rate, what a wrong answer costs, and who would have to own it.

What it covers

  • A ranked shortlist of candidate workflows, with the reasoning visible
  • For each: automate, augment, autonomy — or leave alone, and why
  • The data, state and control implications of each candidate

You end up with: A written assessment you can take to a steering committee — and a defensible reason for the two or three you decide to pursue.

Ask about the AI Workflow Opportunity Scan →

Already building something?

Enterprise AI Architecture Review

The hard part of enterprise AI is rarely the model. It is where the capability sits relative to the systems of record, what it may write and under whose identity, how its actions are logged and reconciled, what happens when an upstream system is unavailable, and what rollback means. This is the review that asks those questions while they are still cheap to answer.

What it covers

  • Placement: read-only, propose-and-queue, scoped write, orchestrating
  • Identity, permissions and who owns which record
  • Controls, segregation of duties, audit trail and observability

You end up with: A written review with the risks named, the gaps prioritised, and the specific changes that would let it run in production.

Ask about the Enterprise AI Architecture Review →

Common questions

What does an enterprise AI workflow redesign engagement cost?

A fixed fee, quoted in the written scope that follows the first call. There is no package price because the scope depends on the workflow and how much of the surrounding landscape is in play — a published figure would fit almost nobody. The first call is free, and if the work turns out smaller than a sprint, I will say so.

How long does an AI workflow redesign sprint take?

A sprint runs in weeks rather than months, and the opportunity scan and architecture review are shorter than that. The shape matters more than the number: if a piece of work cannot be done inside that window it is almost certainly two engagements rather than one long one, and saying so at the start is cheaper for everyone than discovering it halfway. Remote by default, on site where the work genuinely needs it.

Will you sign an NDA before I describe the workflow?

Yes, and before rather than after. Anything you send is treated as confidential, and I am happy to sign your NDA before you describe the workflow — describing the problem is usually the part that involves the sensitive detail, so an agreement that only arrives once the engagement is agreed has been signed too late to protect the conversation that decided it.

What happens if AI is not the right answer for my workflow?

Then I say so, and that is a good outcome rather than a failed one. A scan that comes back with two workflows worth pursuing and honestly none of the rest has done its job — it has saved the cost of the other five. The expensive version is the engagement where nobody is willing to say it, and the programme spends a year proving something that was answerable in a fortnight.

Do you have client references for AI transformation work?

No, and I will not invent them. My relevant background is in enterprise systems — SAP, finance, treasury, and programmes that had to withstand controls, cutover and audit. I am not presenting this as a portfolio of paid AI transformation client engagements, and I have no client logos, testimonials or savings figures to offer. What you can inspect instead is the published method, the practical tools, and the architecture questions I would apply to your workflow. Read a few and you can decide whether the approach fits before we speak.

Get in touch

Tell me about the workflow: what it is, roughly how it runs today, and what makes it painful. If you'd rather read first, everything I'd apply is published — the workflow-design method, the architecture and controls questions, and the guides they come from.

Request a free 30-minute call