The most useful AI case studies are not magic tricks. They show where information came from, which work changed, who reviewed the result, and how the team estimated the benefit. OpenAI's new Stampli customer story contains enough detail to extract a practical operating pattern, provided the headline numbers are treated with care.
Stampli is a procure-to-pay software company connecting procurement, accounts payable, vendors, payments, and employee spending. OpenAI says Stampli used Codex and ChatGPT Work to support the launch of a finance product, linking product context, meeting notes, decisions, and messaging guidance. The company estimated that a defined production workflow fell from 243 active role-hours to about 77, while customer-facing materials retained human review and final approval.
This is a vendor-published customer story based on company estimates, not an independent productivity study. Its value is not the promise that every business will become 3.16 times faster. The value is the workflow design behind the claim.
What happened
According to OpenAI, Stampli moved from prototype to launch in about six weeks. The work included a blog series, emails, a webinar, presentation materials, social creative, a press release, a web page, and sales enablement. The team used Codex to turn evolving product decisions into review-ready assets and to help build an animation before a contractor finished parts of it.
The same underlying system gathered information from product systems and meeting notes to keep documentation current. Agents connected to the company's source of truth then supported content production and helped retrieve business context during meetings.
Three principles stand out. First, the tools were connected to defined business sources rather than asked to improvise from a blank prompt. Second, AI created review-ready work rather than silently owning every final decision. Third, the team described a specific workflow and compared modeled effort, instead of claiming that AI made the entire organization generally “more productive.”
Why it matters for growing businesses
Procure-to-pay and ERP workflows suffer from the same context problem. Information is spread across purchase requests, quotations, purchase orders, receipts, invoices, vendor messages, approval notes, contracts, and accounting records. Employees spend time reconstructing what happened before they can decide what to do next.
AI can reduce that reconstruction effort. It can assemble an invoice packet, compare documents, summarize exceptions, draft a supplier question, or prepare an approval note. But the system must know which source is authoritative. A meeting note does not override an approved purchase order. A supplier email does not prove receipt. An AI-generated explanation does not replace a tax rule or posting control.
This is where ERP design matters. Odoo or another ERP should remain the system of record for approved vendors, orders, receipts, invoices, accounts, taxes, and payments. AI can work across those records, but it should not blur their status or invent missing approvals.
A practical procure-to-pay pilot
- Choose one document-heavy step. Good candidates include invoice intake, three-way-match preparation, vendor onboarding review, or drafting responses to invoice exceptions.
- Map authoritative sources. Identify the ERP records, contract repository, policy documents, and approved communication channels the assistant may use.
- Define the output. Require a structured packet with invoice details, purchase order, receipt evidence, variance, proposed action, and linked sources.
- Make uncertainty visible. Missing receipts, duplicate invoice numbers, tax ambiguity, changed bank details, and amount mismatches must become exceptions, not guessed answers.
- Preserve separation of duties. The person or agent assembling evidence should not also gain unrestricted authority to approve and pay.
- Keep human approval. Vendor creation, bank-detail changes, tax decisions, payment release, and material exceptions should require a named reviewer.
- Measure the real baseline. Track preparation time, review time, exception rate, duplicate detection, correction effort, and days to approval before and during the pilot.
- Audit samples. Review both accepted and rejected recommendations. A system can look accurate while consistently missing one rare but expensive class of error.
What success should look like
A useful pilot produces cleaner review packets faster, reduces the time spent searching across systems, and keeps exceptions visible. It should not merely increase the number of invoices processed if reviewers must later repair classifications or investigate unexplained decisions.
Set a target such as reducing packet-preparation time by 30% while maintaining the existing reconciliation and approval error rate. Run the pilot in shadow mode first. If results are stable, allow the assistant to create draft ERP activities or proposed entries. Posting and payment authority can remain separate until evidence supports a narrower change.
How VDS Consultoria can help
VDS Consultoria helps businesses map procure-to-pay workflows, clean master data, configure Odoo approvals, and add AI where it removes reconstruction work without weakening financial control. We can build document intake, exception queues, source-linked summaries, review steps, and metrics for a controlled pilot.
The Stampli story is encouraging, but the transferable lesson is disciplined: connect AI to the source of truth, define a bounded output, retain final review, and measure a real workflow. Speed becomes valuable when the evidence remains intact. That discipline also makes later expansion easier because every new permission has an owner, a reason, and a measurable control baseline.
Source
OpenAI, “How ChatGPT Work helps Stampli move ideas to market,” 20 August 2026: https://openai.com/index/stampli