Five stages, from Manual to Agentic

Your GAAPx Score shows up as a stage. There are five, numbered 0 to 4, from Manual to Agentic. Each stage covers a band on the 0 to 100 scale. The score is self-assessed, so the stage is too. Nobody requires you to reach Stage 4.

A stage tells you how the work is done today: who does it, how many hand-offs it has and where it waits. It does not grade your team. For how the score is worked out, see how the score works.

Stage, score and readiness

Two scores sit side by side, and they stay separate. The stage is a label on the first one.

• GAAPx Score. How the work runs across the eight areas, from 0 to 100. It is self-assessed.
• Readiness score. Whether your finance team can take on change: talent, operating model, technology, data and leadership alignment. It has its own score, also self-assessed, and never changes the GAAPx Score.

The stage is the label for where your GAAPx Score falls. Stage 2 Automating means a score from 40 to 59.

Two companies can sit at the same stage with different Readiness scores. The one with the higher Readiness score will find the next move easier to carry out.

Stage	Score band	In one line
Stage 0 Manual	0 to 19	Most work runs on people, email and spreadsheets.
Stage 1 Digitized	20 to 39	Data arrives electronically, and people still carry it between systems.
Stage 2 Automating	40 to 59	Rules handle the main path, and people handle the exceptions.
Stage 3 Agent-assisted	60 to 79	Agents prepare the work and a person decides.
Stage 4 Agentic	80 to 100	Agents and code do the routine work inside tested controls, a named person owns each control, and people handle exceptions and decisions.

Beyond Stage 4 is not scored.

Every example below uses the same two processes, bank reconciliation and supplier invoices, so you can see how the same work changes from stage to stage.

Score bands

Stage	GAAPx Score
Stage 0 Manual	0 to 19
Stage 1 Digitized	20 to 39
Stage 2 Automating	40 to 59
Stage 3 Agent-assisted	60 to 79
Stage 4 Agentic	80 to 100

These bands are our starting assumption. They split the 0 to 100 scale into five equal parts. We have not checked them against other companies, and we will review them as we learn from real use.

The five stages side by side

	Process shape	Who does the work	Where waits sit	Typical step mix
Stage 0 Manual	One step after another, passed by email or paper. Ownership is informal.	People, by hand	In inboxes and with whoever has time	Human steps, many of them Delete candidates such as keying
Stage 1 Digitized	Same steps, now in software. Hand-offs and queues stay.	People, with data already loaded	Between systems and at each export and import	Human steps, with Delete candidates still open
Stage 2 Automating	Fewer hand-offs, an exceptions queue, one named owner.	Rules and code on the main path, people on exceptions	In the exceptions queue	Plain code leads, Human decisions on exceptions
Stage 3 Agent-assisted	The flow has agent steps that prepare work for a person.	Agents prepare, a person decides each item	Waiting for a person to review the agents' suggestions	Plain code and Agents, Human decisions on each item
Stage 4 Agentic	Rebuilt around agents. Few hand-offs, one accountable owner, a named person for each control.	Agents and code do routine work, people handle exceptions and decisions	On flagged exceptions and owner review	Plain code and Agents lead, Human decisions on what is flagged

What a stage is not

A stage is not a grade for your team. A small team with a clean Stage 1 close is in a better place than a large team with a messy Stage 2 one.

A stage is not a standard. Nobody requires you to reach Stage 4.

A stage is not a peer comparison. We compare you to no other company.

A stage belongs to one area at a time. Your team can sit at Stage 3 in Cash Ops and Stage 1 in Spend.

The four step buckets

Each stage below names the step buckets that dominate it. Every step in a workflow falls into one of four.

• Delete. The step need not be done at all, or it disappears when data is captured once.
• Plain code. A fixed rule or script does it the same way each time.
• Agents. An agent does the step, including a fixed flow with an agent step in it.
• Human decisions. A person has to judge, approve or answer for it.

Stage 0 Manual

Definition. Most transactions are handled by hand, including keying data into systems.

How the work is shaped. Work moves one step after another, passed along by email, paper or a shared folder. Each hand-off waits for the next person to pick it up. Nobody owns the process from end to end, so whoever has time carries it forward. Status lives in a spreadsheet or in someone's head.

Who does the work. People do every step. Rules and code do almost nothing. There are no agents. The approver decides, usually by replying to an email, and the decision is hard to find later.

What you typically see

• Someone types bank lines or invoice details into the ledger.
• Approvals happen by email or on paper, and the trail is an inbox.
• Status lives in a spreadsheet that one person keeps up to date.
• Close tasks wait on whoever has time.
• Vendor bank details arrive by email and are copied by hand.
• When one person is away, a task stops.

Waits at this stage. Unplanned waits are the main story. Work is finished, and the next step has not started because it sits in an inbox or waits for a person to find time. A coded invoice that sits with an approver for four days is a typical case. Planned holds exist here too, such as paying a supplier on day 30 by design. You choose those, and they are not counted against you.

How to tell you are here

1. Does someone type invoice or bank data into your ledger by hand most days?
2. Do approvals run through email or paper?
3. Would a task stall for days if one named person were out?

Three yes answers put you at Stage 0.

Example. An accounts payable clerk receives supplier invoices as PDFs by email. The clerk keys each one into the ERP, emails the approver and chases for a reply. At month end, the accountant downloads the bank statement, prints or pastes it next to the ledger and ticks off matches line by line.

What usually pays off next. Cut the repeat work. Switch on bank feeds and take invoice entry out of the inbox. Both remove keying without changing how your team makes decisions. To reach Stage 1, get data into your systems electronically and put the status in one shared place.

Common traps. Buying an agent tool while invoices still arrive as email attachments. The tool has nothing clean to work on. Another trap is counting a spreadsheet tracker as a system. It helps, and it still depends on one person.

Step buckets. Human steps dominate, and many of them are Delete candidates, such as keying from an email.

Stage 1 Digitized

Definition. Data arrives electronically, and people still carry the work between systems.

How the work is shaped. The steps are the same as at Stage 0, now running in software. Bank feeds or an ERP link bring data in. The hand-offs and queues stay. Work still moves when a person exports from one system and imports into another. Someone owns each task, and nobody owns the whole flow.

Who does the work. People do the matching, coding and chasing, with the data already in front of them. Rules and code load and format the data. There are no agents. People decide every item.

What you typically see

• Bank transactions load into the ledger each day.
• Standard templates and shared checklists exist for the close.
• Someone exports from one system and imports into another.
• Matching and coding are done by a person, one line at a time.
• Invoices are read from images into fields, and a person checks each one.
• A spreadsheet of open items goes to the controller each week.

Waits at this stage. Unplanned waits move to the gaps between systems. A file sits until someone exports it. An invoice is captured and then waits for a person to route it. A reconciliation is done and the reviewer has not started. Planned holds look the same as at Stage 0, for example the day a subledger locks by design.

How to tell you are here

1. Do bank lines and invoice data reach your ledger without anyone typing them?
2. Does a person still match, code and route each item?
3. Does someone still export from one system and import into another each period?

Yes, yes and yes put you at Stage 1.

Example. Bank transactions flow into the ERP through a feed. The accountant still matches them to ledger entries one at a time and exports a spreadsheet of open items for the controller. Supplier invoices are captured as images and read into fields. A person checks each one and routes it for approval.

What usually pays off next. Let the tools run without someone pushing every step. Set matching rules for the lines that repeat every month, and route approvals by amount and supplier. To reach Stage 2, remove hand-offs before you add rules, then let rules run the main path.

Common traps. The automated mess. Putting software on the same steps keeps the same queues and amplifies what is wrong. If an invoice passes through five hands today, a tool that speeds up each hand still leaves five. Remove hand-offs first, then automate. Ask of each step whether it needs to happen at all.

Step buckets. Human steps still dominate. Delete candidates remain, such as the export and re-import between systems. Plain code starts to appear for loading and formatting data.

Stage 2 Automating

Definition. Rules handle the main path, and people handle the exceptions.

How the work is shaped. The path has fewer hand-offs. Steps that can run side by side do. Items that fit the rules go straight through. Items that do not fit land in an exceptions queue with an owner and an age. One named person owns the process from start to finish.

Who does the work. Rules and code do the routine matching, coding and routing. People work the exceptions and keep the rules current. There are no agents yet, or they are a side experiment. A person decides every exception, and a rule decides the clear cases.

What you typically see

• Most bank lines match on their own, and a short list of unmatched items goes to a person.
• Invoices that match the purchase order and receipt post without a touch.
• Exceptions show up in a queue with an owner and an age.
• The rules are written down, and someone reviews them when something changes.
• Approval routing follows amount and supplier, and nobody forwards items by hand.
• Close tasks start on their own when the one before finishes.

Waits at this stage. Unplanned waits gather in the exceptions queue. An item flagged on Monday and untouched on Thursday is an unplanned wait. The queue's age column makes it visible. Planned holds are easier to see now because the rules can carry them, for example a payment scheduled for day 30 on purpose.

How to tell you are here

1. Does most of your volume go through without a person touching it?
2. Does every exception land in a queue with an owner and an age?
3. Can you show the written rules to an auditor?

Yes, yes and yes put you at Stage 2.

Example. The reconciliation tool matches recurring receipts and payments by rule. The accountant works only the unmatched list. Supplier invoices that agree with the purchase order and receipt go straight to payment approval. Those that do not land in an exceptions queue.

What usually pays off next. Use agents on the exceptions and on the judgment work. Suggested matches for the unmatched list and suggested coding for odd invoices are good places to start. To reach Stage 3, pick one queue, let an agent prepare each item with its reason, and keep a person deciding.

Common traps. Reaching Stage 2 by adding rules to a Stage 1 flow without removing hand-offs first. That is the automated mess again, and the exceptions queue fills with items the old steps created. Another trap is rules that nobody reviews. They drift when suppliers, banks or chart of accounts change.

Step buckets. Plain code dominates, since deterministic rules run the main path. Human decisions cover the exceptions.

Stage 3 Agent-assisted

Definition. Agents prepare the work and a person decides.

How the work is shaped. The flow has agent steps inside it. Agents read, propose and flag. A person reviews each proposal before anything posts. The exceptions queue is shorter because agents resolve some items into suggestions. One owner is accountable for the process and for the review step.

Who does the work. Rules and code handle the main path. Agents prepare matches, codes, accruals and explanations. People review, accept or change each suggestion, and a person decides every item. Your team knows which suggestions it accepts and which it overrides.

What you typically see

• Agents suggest matches, codes or accruals, and your team accepts or changes them.
• Each suggestion shows its reason or its source.
• People review the suggestions before anything posts.
• Your team can tell you how often it accepts or overrides the suggestions.
• Someone has written down which data the agents may use.
• Duplicate invoice flags arrive before payment, not after.

Waits at this stage. Unplanned waits shift to review. A suggestion that is ready and sits for two days before a person looks at it is an unplanned wait. Software can prepare an item in seconds, so review time often becomes the largest wait. Planned holds stay as chosen, such as payment on the due date.

How to tell you are here

1. Do agents propose matches, codes or accruals in your live work today?
2. Does a person review each proposal before it posts?
3. Can your team say how often it accepts an agent's suggestion?

Yes, yes and yes put you at Stage 3.

Example. For bank reconciliation, an agent proposes a match for each unmatched line and explains why. The accountant approves or corrects each proposal. For supplier invoices, an agent suggests the coding and flags a possible duplicate, and the approver confirms.

What usually pays off next. Stronger review and controls. Record who approved what, test that the checks catch mistakes, and set rules for which suggestions need a second pair of eyes. To reach Stage 4, move the clear, low-risk items from "a person approves each one" to "a person reviews a selection and the exceptions", inside controls you have tested.

Common traps. Approving suggestions without reading them. The review step exists on paper while the habit is to click accept. Another trap is leaving the review queue unowned, so agents speed up preparation while the work waits for a person.

Step buckets. Plain code and Agents share the routine work, and Agents here includes a fixed flow with an agent step. Human decisions sit on each item.

AI oversight. This is our own four-level scale of who checks the AI's work. It is separate from the stage and never changes your score.

Stage 4 Agentic

Definition. The work is rebuilt around agents. Agents and code do the routine work inside tested controls, a named person owns each control, and people handle exceptions and decisions.

How the work is shaped. The flow has few hand-offs. Routine items go from start to finish without a person in the middle. People review a selection of items, the flagged exceptions and a daily summary. One named owner is accountable for each process where agents do the work. Monitoring watches for drift, errors and unusual activity.

Who does the work. Agents and code complete routine tasks end to end. People handle exceptions, set the limits and decide what the work should do. The owner answers for the result. Controls are tested on a schedule, and the evidence is kept for audit.

What you typically see

• Agents complete routine tasks end to end, and a person reviews a selection of items and the exceptions.
• Controls are tested on a schedule, and the evidence is on hand for audit.
• Monitoring flags drift, errors and unusual activity.
• A named owner is accountable for each process where agents do the work, and a named person owns each control.
• Limits are set per supplier and amount, and anything above a limit goes to a person.
• The controller starts the morning with a summary and a short list of exceptions.

Waits at this stage. Few unplanned waits remain, and they sit where a person has to act: a flagged exception, an owner review, a limit that needs approval. Planned holds are built into the flow on purpose, such as payment on day 30, and are shown and left out of the score.

How to tell you are here

1. Do routine items finish without a person touching each one?
2. Do you test your controls on a schedule and keep the evidence?
3. Is there a named owner for each process where agents do the work, and a named person for each control?

Yes, yes and yes put you at Stage 4.

Example. An agent reconciles the bank accounts each day and posts the matches. The controller reviews a summary and the exceptions each morning. Supplier invoices within set limits are coded, matched and queued for payment, with an audit trail for each step.

What usually pays off next. Keep the controls proven. Retest them when systems, volumes or suppliers change, and keep the evidence current. Stage 4 is the top of the scored ladder, so the work from here is to keep the result and keep watching it.

Common traps. Treating Stage 4 as finished. Controls that passed a test last year may not fit this year's volumes. Another trap is removing review to look faster. Segregation of duties and a person's sign-off on limits still matter.

Step buckets. Plain code and Agents do the routine work. Human decisions remain for what the controls flag and for the limits themselves.

AI oversight. This is our own four-level scale of who checks the AI's work. It is separate from the stage and never changes your score.

Beyond Stage 4

The fully agentic enterprise is not here yet. Beyond Stage 4 has no name and no score, and nothing on this page counts toward it.

How stages relate to the 0 to 10 scale

Each process is rated 0 to 10, and the journey score is shown from 0 to 100. The anchors on the journey scale sit inside the bands above. Rating 0 is Stage 0 Manual. Rating 2 is the start of Stage 1 Digitized. Rating 4 is the start of Stage 2 Automating. Rating 6 is the start of Stage 3 Agent-assisted. Rating 8 is the start of Stage 4 Agentic.

See how the score works for the separate efficiency measure and the GAAPx Score.

Questions people ask

Can one area be Stage 4 while another is Stage 1? Yes. Stages belong to an area, not to the whole company. Cash Ops can run on agents and code while Spend still relies on email approvals.

Why these score bands? Five equal parts of a 0 to 100 scale are the simplest split. They are our starting assumption, and we have not checked them against other companies. We will review them as we learn from real use.

What if I am between stages? Take the lower one for now. A stage describes what you see most of the time. If your checks split, for example two yes answers at Stage 3, the next step is the missing check. That check is your next piece of work.

Does anything beyond Stage 4 count? No. It has no name and no score yet. It does not change your GAAPx Score or your Readiness score.

Does a higher stage mean a better team? No. A stage describes how the work is done. It says nothing about the people doing it.