Skip to content
CodelessOps
Go back

The seams map

Karim Lameer

Karim Lameer — Master Anaplanner, CIMA-qualified, 15 years in finance and FP&A. I build Grounded and published the Board Pack Test. LinkedIn · New here? Start here

It’s day three of the close and the headcount file hasn’t arrived. Someone in HR is exporting it. Someone in finance will tidy the columns, load it, find that it doesn’t tie, and ask for it again.

Nothing is wrong with the HR system. Nothing is wrong with the planning model. The morning goes on the gap between them.

I call that gap a seam: any point where a number leaves one system and a person carries it into another. An export, a re-key, a file on an email, two windows open side by side. Once each system is set up well, that is where the hours go, and where the errors get in.

By the end of this post you’ll have your seams on one page: every system you run, every hand-off a person does by hand, who does it, how often, and what broke last time. From that page you’ll name the first seam worth fixing and sort the fix into one of three jobs: a pipe can do it, something has to read evidence, or a person decides. In the one close I’ve written up, it was a pipe, not AI.

Half of FP&A forecast time still goes on gathering and preparing data. That was the share in 2020, and the same survey said it was the share a decade earlier.1 Ten years of new systems, and the hours didn’t move.

Where the month goes

The surveys that ask what slows the close keep naming the same things, and none of them is a system. In Sage Intacct’s 2021 survey of 1,600 finance people, the closes most hit by manual imports and exports ran 25% longer than the rest, and 58% said information hand-offs had a medium or high impact on their close time.2 In the UK, 86% of finance decision makers at medium-sized organisations now say the close takes more than three days, up from 72% a year earlier.3

Two honest labels. Every survey in this post was run by a vendor with something to sell, and every one measures gathering or delay in general, not time between systems.45678 I read them as a pattern, not a measurement. The pattern is enough: nothing in that list is a system that’s broken.

Three questions before you draw the map

Before the map, a check. The biggest documented wins I could find were inside one system, not between two. An education centre in Canada cut its chart of accounts from 1,800 lines to about 400 and its close went from three weeks to one; the implementer tells the story, so read it as such.9 “Fix the data before layering on AI” is the line the trade press keeps landing on.10

I agree with it. A map of hand-offs around a messy ledger will send you to automate a mess. So three yes-or-no questions first. I made these up from my own builds, since nobody publishes a gate for this.

Before you draw the map
  1. Is there one chart of accounts that every entity uses, with a named owner?
  2. Does every master list (suppliers, customers, cost centres) have one owner, and no known duplicates?
  3. Do the loads you already have run without a person touching them?

Two “no”s and the map waits a quarter. Fix the inside, then draw. In my own close, two of the five rows were inside fixes: consolidation moved from linked workbooks into the model, and the team stopped leaving the model to review numbers. Neither was a hand-off automated. Both were a boundary removed.11

The map is for a finance function whose boxes already work.

The map: one page, nine columns

Here it is. One row per seam. Three rows filled so you can see the shape; the rest are yours.

FromToBy hand, howWhoHow oftenHours per closeWhat broke last time, and what it costsHow often either end changesWhose seam
HR systemPlanning modelexport, tidy the columns, uploadFP&A analystevery close4a renamed column; the plan ran on last month's headcountquarterly (new department codes)unreached
Bank statementLedgertwo windows, matched by eyeassistant accountantweekly6a receipt matched to the wrong invoice; a customer chased who had paidrarelyreached (check the demo)
Planning modelBudget holder's inboxpaste into an email, chase for a replyfinance business partnerevery close3a late reply; the accrual guessedwhen people movea person
The seams map. One row per hand-off. The three rows are examples of the shape, not of what yours will say; the hours are illustrative.

The columns, left to right. From and to: a system, or a person (the budget holder’s inbox is a seam too). How it’s done by hand: export, re-key, email, paste, or two windows side by side. Who does it, and how often. Hours per close: rough is fine, the person doing it knows.

Then the two columns that feed the score. What broke last time, and what a wrong number here costs: a restated pack, a late filing, a supplier paid twice. How often either end changes: a new cost centre, a renamed column, a supplier set up again under a different name. The last column, whose seam it is, gets its own section next.

How to fill it: one person, one hour, with the people who do the hand-offs in the room. No vendor present. Most teams find more lines than boxes. That’s normal. The month runs on the lines.

Plenty of worksheets exist for choosing what to automate. The ones I found either score a process and never the join between two systems, or ask good questions about a hand-off and never put a number on it.121314 This one puts the numbers on the join.

Whose seam is it

Most of the vendors you run have shipped an assistant or an agent in the last year or two, and every one I’ve read works on the data inside its own product. Microsoft’s reconciliation agent in Dynamics 365 reconciles the ledger “with the accounts payable, accounts receivable, tax, and bank subledgers”.15 Workday’s finance agents are “powered by Workday’s unmatched data and context”.16 NetSuite, Xero and Sage say the same in their own words.171819

Two qualifications, because the picture moved this year. Oracle’s Payables Agent “can ingest invoices from email, portals, EDI/e-invoicing, and PDFs”, and its cash agent reads bank statement lines.20 SAP says its prebuilt Joule Agents “reliably execute multistep tasks across SAP and non-SAP systems”, though the page names no finance agent that does.21 What every vendor offers for the rest is a studio: a place where your team builds the connection to the other product, and then keeps it working.222324

So the last column takes one of three values.

Reached: the seam ends in a product whose vendor ships an agent for that hand-off. Invoices arriving by email into the ledger. The bank statement into the ledger. Check the demo, because it may already cover the row.

Unreached: the seam runs between two products. The HR system into the planning model, the CRM into the ledger, the ledger into the pack. Nobody’s, unless you build it. (One consultancy draws the same line: vendor agents sit on workflows confined to a single system, and cross-system work gets built in-house.25)

A person: the seam ends in someone’s inbox. The sort further down still applies.

Bank CRM HR system Ledger Planning model The pack Budget holder reached by the vendor's agent unreached: between two products ends in a person
A typical estate, the seams marked by whose they are. One reached, five unreached, one that ends in a person. Yours will differ; count them.

Count the three kinds on your page. The demo shows you the seams that end in the product. The ones between products are yours.

What breaks, loudly and silently

The eighth column asks how often either end changes, and it’s there because of what happens when one does. A column gets renamed upstream, a code gets added, a supplier gets set up again. The pipe between the two systems then does one of two things.

It stops: the load fails and someone fixes it in the morning.2627 Or it forks: the renamed field becomes a new column and the old one quietly fills with nothing.28 Nothing I read decides a rename is the same field. A person decides that, or nobody does.

The silent kind has a famous case. In October 2020 Public Health England lost 15,841 COVID cases from the daily count because lab results were loaded into an old Excel format with a 65,536-row cap; the rows past it “were simply left off and not counted when imported”.29 And hand-keyed cells carry their own error rate before any of that.30

The column asks how often either end changes, and who hears when it does.

Pick the first seam

Now score the rows. I use four factors, in order, and I made the rule up, since nobody publishes one for seams.

  1. Hours per close times runs a year. The size of the prize, in time released.
  2. Cost of the error. What a wrong number here has cost, or would.
  3. Stability. A grade off for each end that changes often.
  4. An owner. An unowned seam loses a grade, and finding its owner is the first fix.

The highest row goes first. “Pays” means time released per close and errors stopped. Never cash saved. I don’t put cash on a seam before it has run through a close.

Two traps. The first is the seam that’s really a judgement call: the accrual that waits for the project lead’s view. Automating the chase helps. Automating the decision doesn’t, and the row should say which it is.

The second is the seam whose upstream changes every month. Build the validation first and the pipe second, or the pipe will be the thing that breaks.

A third I’ve met: the seam nobody can explain. A workaround meant for one quarter, still running, its reason gone with whoever built it.31 Score it, but ask why it exists before you automate it.

The rows you don’t pick wait for next quarter’s map. The row at the top of the page is the brief for whoever builds it, in-house or not.

Sort the fix: a pipe, something that reads evidence, or a person

Take the top row and walk it step by step. Each step is one of three jobs. I wrote this sort for the work around planning models, and it holds for any seam.31

A rule or a pipe can do it: the file arrives on schedule, the columns are the ones expected, the totals agree. The export, the email and the upload themselves.

Something has to read evidence: a column has been renamed, and is it the same thing? The note from HR says why. Two files side by side, the differences listed, a fix proposed with what’s still uncertain.

A person decides: approve the change, sign the rerun, own the mapping.

THE HEADCOUNT FILE, STEP BY STEP THE JOB 1 · HR exports the file 2 · The file reaches finance 3 · Columns checked against the ones expected 4 · "Department" is now "Supervisory organisation": same thing? 5 · HR's change note read; the two files compared 6 · The mapping change approved 7 · Load, and the totals tie PIPE PIPE PIPE READS EVIDENCE READS EVIDENCE PERSON PIPE Four steps a pipe can do, two that need something to read, one that stays with a person.
The headcount file from the opening, as seven steps. Most of a seam is pipe work.

That’s the headcount file from the opening, as steps. Four are pipe work, two need something that reads, one is a person. In the one close I’ve published, the list of what moved it had no AI in it at all.11 Where a vendor already ships an agent for the seam (the reached rows), the first fix may be that agent.

The vendors’ own criteria agree with the sort, by the way: stable underlying applications first, and never a process that keeps changing.32 An integration vendor puts it as: make each hand-off deterministic first, then give agents access.14

The sort is the same for every row. Only the proportions change.

Twelve days to eight hours, as rows

The method, on the one close I’ve written up: a listed, clinical-stage biotech, the mechanical close, ledger close to consolidated actuals ready for review.11 Five rows. Per-row hours were never measured; only the whole thing was. From twelve working days down to under eight hours, and the reforecast from a week to one day.

FromToBy hand, beforeWhat replaced itSort
NetSuite (actuals)Anaplanexport requested, columns tidied, file loaded; requested again when it didn't tiescheduled overnight loadspipe
Two HR systems (headcount)Anaplanon its own separate loopscheduled loadspipe
Snowflake (warehouse data)Anaplanloaded by handscheduled loadpipe
Ledger extractsConsolidation workbookslinked workbooks that waited for everything elseconsolidation moved inside the modelboundary removed
The modelReview outside itnumbers left the model to be revieweda complete UX; the team stays inside the modelboundary removed
What stayed with people: accruals, cut-off calls, review and explanationperson
One close on the map. Three pipes, two boundaries removed, the judgement left alone. No AI anywhere in it.

The write-up lists the loads first and the consolidation move second, and I’d keep that order. It doesn’t say which came first, so I won’t claim it here.

One more row, from a different seam. At Fox, every supplier invoice was keyed and coded by hand before it could go for approval, more than 60 hours a month. Now the invoices are read, coded and queued as they land, and the approval stays with a person. About two hours a month.33

That one wasn’t a pipe. It was something that reads evidence, with a person at the end. And on today’s map it sits in a reached row, because the ledger vendors now ship that agent themselves.

Move the numbers first. Once they move themselves, put the intelligence on top.

What I’d tell your CFO

Map before you buy. One page, one hour, the people who do the hand-offs in the room, before the next vendor call. The demo will cover the seams that end in its product; mark those reached and check them. The seams between your products are yours, and in my close those were pipes, scheduled loads and a boundary removed, with no AI budget. If the question is “why not just buy the ERP vendor’s agent”, buy it where the seam ends in their product, and draw the rest, because nobody else will.

If you’d rather I drew it, the scoping sprint on the services page ranks the seams and scopes the first build.

Sources

Footnotes

  1. AFP, Survey: forecasting services to remain in demand in 2021, press release, 8 Dec 2020. 484 FP&A practitioners; underwritten by Workday; “half its forecast time in data gathering and data preparation”. ↩

  2. Kellie Pharr, The top 5 obstacles slowing down your close, Sage Intacct, 21 Jul 2021. Close the Books survey, 1,600 finance professionals; vendor-run. ↩

  3. Darren Slade, How you can achieve a stress-free month-end close at last, iplicit, 19 Oct 2025. 1,000 finance decision makers, medium-sized UK organisations; commissioned by iplicit. ↩

  4. FP&A Trends Group, 2025 FP&A benchmarks and trends, 10 Jul 2025. 46% of FP&A time on data collection and validation; sponsored by OneStream. ↩

  5. Alexei Alexis, CFOs push for a faster month-end close, CFO Dive, 5 May 2026. LiveFlow survey; nearly eight in ten blame waiting on other systems; sample not reported. ↩

  6. Ledge, Month-end close benchmarks for 2025, Apr 2025. 100 finance professionals; cash reconciliation pulls from three to five systems; vendor-run. ↩

  7. Journal of Accountancy, Finance leaders have trust issues with their data, 31 Jan 2024. BlackLine survey, 1,300+ in seven countries; 31% cite too many data sources. ↩

  8. IT Brief UK, UK CFOs still rely on spreadsheets for financial close, 18 May 2026. Sixthfin study by Odoxa, 303 UK CFOs; 67% use Excel for reconciliation. ↩

  9. MNP Digital, Kahnawà:ke Education Center, undated. Chart of accounts 1,800 to about 400 lines; close three weeks to one; implementer’s case study. ↩

  10. ERP Today, Finance transformation stalls where data assembly begins, Sep 2026. On Unit4’s finance-data piece; “fix the data before layering on AI”. ↩

  11. Karim Lameer, 12 days to 8 hours: what actually moved the close, codelessops.com, 11 Sep 2026. The full list of what moved the Exscientia close; no AI in the chain. ↩ ↩2 ↩3

  12. KeepSolid, Map cross-functional handoffs before automating, 21 Jul 2026. Seven questions per hand-off; owners and evidence, no numbers or score. ↩

  13. Centric Consulting, RPA scorecard, Apr 2023. Scores one process on eight sliders; never the join between two systems. ↩

  14. Adam White, Order to cash process, Tray.ai, 2 Oct 2026. Five hand-offs between systems; “make each handoff deterministic first, then give agents access”. ↩ ↩2

  15. Microsoft Learn, Account reconciliation, Dynamics 365 Finance, 21 Jul 2026. Ledger reconciled with its own subledgers; no external system named. ↩

  16. Workday, Workday Illuminate expands with new AI agents, press release, 16 Sep 2025. Finance agents “powered by Workday’s unmatched data and context”. ↩

  17. NetSuite, NetSuite 2024.1 brings AI to life, 2 Aug 2024. Text Enhance drafts “using relevant NetSuite data”; Bill Capture matches invoices inside NetSuite. ↩

  18. Xero, JAX, undated. “Using your unique business data in Xero and real-time public information.” ↩

  19. Sage, Sage Copilot, undated, read 19 Oct 2026. “Embedded within the Sage Platform”; features marked illustrative. ↩

  20. Oracle, Oracle AI for Enterprise Resource Planning, undated, as read 19 Oct 2026. Payables Agent ingests invoices from email and portals; Cash Processing Agent reads bank statement lines. ↩

  21. SAP, Joule Agents and Joule Assistants, undated, read 19 Oct 2026. “Across SAP and non-SAP systems” via MCP and A2A; no finance agent named on the page. ↩

  22. Salesforce, How Data Cloud powers Agentforce, 20 Nov 2024. Agents grounded in Data Cloud; outside data via zero copy or a connector the customer builds. ↩

  23. SAP, Joule Studio, undated. Organisations create their own agents “across SAP and non-SAP systems”. ↩

  24. Microsoft Learn, Connectors overview, undated. Prebuilt connectors for common cases; custom connectors built by the customer. ↩

  25. Bain, Is agentic AI the inflection point for scaling ERP transformations?, undated. Vendor-partnered agents on single-system workflows; cross-system built in-house. ↩

  26. Workato docs, Salesforce: troubleshoot runtime errors, undated. A recipe still requesting a deleted field fails until the schema is refreshed. ↩

  27. Anaplan Community, Import unsuccessful: ensure the column headers in the import file, undated. Vendor-run forum; an import fails when the file’s headers change since it was built. ↩

  28. Fivetran docs, Custom field name changes, undated. A renamed field becomes a new column; the old one fills with nulls. ↩

  29. Jude Karabus, Excel England coronavirus contact error, The Register, 5 Oct 2020. 15,841 cases; 65,536-row cap; “simply left off and not counted when imported”. ↩

  30. Raymond Panko, Spreadsheet errors: what we know, what we think we can do, EUSPRIG 2000 (arXiv 2008). Errors in 51% of 1,170 lab spreadsheets; cell error rates 1.1% to 2.5% in careful audits. ↩

  31. Karim Lameer, Anaplan has the number. Finance still has the work., codelessops.com, 13 Sep 2026. The three-way sort; the exclusion nobody can explain. ↩ ↩2

  32. UiPath, Identifying the right processes for automation, undated. “Stable processes and underlying applications”; never a process constantly going through changes. ↩

  33. Karim Lameer, Invoice processing, from 60-plus hours a month to about two, codelessops.com. Invoices read, coded and queued as they land; approval stays with a person. ↩


Want the manual work around your close taken out?See how I build it: scoped first, built in phases at a fixed price, with a named person approving every output.

Share this post:

Keep reading

All posts →