Start with the task and choose the right tool

Many improvements labelled AI are established automation. Approval routing, scheduled reminders, electronic data interchange and matching within agreed tolerances can work through defined rules. Document capture may use conventional optical character recognition, AI-assisted extraction or a combination. The label alone tells a buyer very little.

AI becomes useful where a task involves varied language, less structured information or patterns that are difficult to express as fixed rules. A model might suggest a spend category from a free-text description, summarise a supplier response or help an analyst explore unusual transactions. Those suggestions still need suitable validation.

This distinction matters commercially. A broken approval route may need a configuration change. Repeated invoice exceptions may need better purchase orders. Introducing AI before understanding the cause can add cost and another layer to maintain.

Start by observing the work, including spreadsheets, email exchanges and informal workarounds. Then decide whether the best response is a clearer process, better data, conventional automation or AI assistance.

01 / source

Source with a clearer evidence base

Sourcing teams often spend considerable effort assembling information before they can make a commercial decision. AI can help organise approved research, turn a business requirement into draft supplier questions and summarise responses against a common structure.

For example, a buyer comparing facilities-management proposals could use an approved tool to extract coverage, mobilisation assumptions and exclusions into a working comparison. The buyer would check each point against the original response and ask suppliers to clarify gaps. Time saved on transcription could be spent testing whether the proposed service will actually meet the business need.

The evidence needs to remain visible. A summary should link back to its source, distinguish a supplier’s claim from verified information and show where no answer was found. Risk and sustainability information also needs a date and a credible source; an AI-generated score is not a substitute for due diligence.

Procurement retains ownership of the criteria, supplier engagement and recommendation. Stakeholders still decide the acceptable trade-offs between price, service, resilience and risk.

02 / contract

Contract with faster preparation and review

Contract work contains many tasks where a prepared first pass can help. With access to approved templates and a maintained clause library, AI can assist with drafting, compare wording with a preferred position and extract obligations for review.

Consider a procurement manager reviewing a supplier’s proposed agreement. A tool could flag differences in payment terms, renewal provisions or service commitments and point the reviewer to the relevant passages. A useful output would explain what changed and show the text it relied on, rather than simply label the agreement acceptable.

Legal and commercial owners then assess the consequences in context. A clause that suits one purchase may be unsuitable for another. Summaries can also miss qualifications, schedules or interactions between terms, so material decisions require checking the underlying agreement.

After signature, reviewed dates and obligations can feed a contract register. Routine reminders may then be handled through ordinary workflow. Someone must still own the renewal decision, notice requirements and follow-up with the supplier.

03 / buy

Buy with guidance at the point of need

Employees generally want to obtain what they need without becoming procurement-system experts. A well-designed buying journey can help them describe a requirement, find the appropriate catalogue or framework and provide the information an approver needs.

AI can support that journey by interpreting a request and suggesting the right next step. An employee asking for a new software tool might be guided towards an existing agreement or prompted to explain the intended use before the request reaches IT and procurement.

That guidance needs to reflect current policy, approved suppliers, catalogues and the requester’s permissions. If the underlying content is outdated, a conversational interface can make the wrong answer easier to obtain.

Budget checks, approval limits and purchase-order controls should remain in the authorised workflow. The system should make clear when it has offered advice, prepared a requisition or actually submitted something. Buyers and budget owners retain the decisions assigned to them, with a clear route for unusual requirements.

04 / receive

Receive with better information about exceptions

Receiving is where a tidy purchasing record meets what actually happened. Goods may arrive in part, the quantity may be wrong, or a service milestone may need a business owner’s acceptance. These details affect stock records, invoice matching and supplier payment.

Electronic delivery messages and supplier portals can reduce rekeying without requiring AI. Where the tools support it, AI may help interpret delivery documents, identify possible inconsistencies or prioritise orders that need attention.

For example, a delivery note could indicate that only part of a purchase order has arrived. A system might prepare the relevant information for the receiving team and highlight the remaining quantity. The authorised receiver still confirms the physical delivery. For services, the accountable owner confirms that the agreed work or milestone has been completed.

A supplier’s delivery message should not automatically become proof of acceptance. Clear ownership, timely recording and a sensible exception route are essential; otherwise the same issue will reappear in accounts payable.

05 / invoice

Invoice with less rekeying and better routing

Invoice processing is an obvious place to look for repetitive work. Depending on the platform, AI-assisted extraction can help capture information from varied invoice layouts, while matching rules check it against purchase orders and receipts. Other models may help identify unusual patterns or suggest where an exception belongs.

Take an invoice that fails to match because a receipt is missing. An effective process identifies the responsible receiver and presents the purchase order, invoice and outstanding action together. Asking accounts payable to investigate the same gap repeatedly wastes the benefit of faster extraction.

Confidence thresholds can help determine when extracted information needs review, but a confident-looking answer can still be wrong. Test results and the consequence of an error should determine the controls. Supplier identity, invoice reference, amounts, currency and tax treatment need appropriate validation. Duplicate alerts or unusual payment requests are prompts to investigate, not proof of fraud.

The goal is fewer avoidable exceptions and less effort per genuine exception. Recording why invoices fail helps the team identify problems upstream, such as unclear ordering instructions, inconsistent supplier records or late receipts.

06 / pay

Pay with stronger preparation and control

At the payment stage, assistance can help finance teams assemble a clearer picture of invoices due, unresolved holds and potential early-payment discounts. Forecasting models may also support cash planning where the available data and implementation are suitable.

The commercial decision needs context. A discount may be unattractive once liquidity needs are considered. A due date may be subject to a dispute. An unusual transaction may be legitimate. Finance and treasury retain responsibility for payment policy, cash decisions and release authority.

Supplier bank-detail changes deserve particular care. A persuasive email or an apparently plausible AI summary must never replace the organisation’s independent verification process. Changes should be checked using established, trusted contact details and authorised controls, with segregation between maintaining supplier data and approving payments where required.

AI can prepare information and flag items for attention. Payment approval and execution should stay within controlled systems, with appropriate access, recorded approvals and a traceable audit trail.

Make the foundations strong enough to support the change

Across all six stages, the same practical dependencies appear. Data, people, processes and technology determine whether assistance reduces work in practice.

The practical expertise behind better outcomes

Start with one measurable pilot

A useful first pilot has a bounded task, enough repeated volume to evaluate and a named process owner. Choose a problem the team already recognises. Invoice-exception triage, for example, could test whether the system can suggest the right reason and owner while the existing approval and payment controls remain in place.

Start with reviewed historical examples and a separate test set. Include partial receipts, missing purchase orders, duplicate-looking invoices and unfamiliar descriptions. Agree what constitutes a correct suggestion, which cases must go straight to a person and when the pilot should be stopped or adjusted.

Run the initial live stage in review mode, with people checking suggestions before any action. Keep access limited and give users a straightforward way to report errors. Do not assume the system will safely learn from corrections on its own; changes need an approved improvement and testing process.

Measure more than the number of transactions processed:

• Hands-on time per case, including checking and correcting the output
• Correct classification and routing, alongside missed and false alerts
• Time taken to resolve an exception and the amount of repeat work
• Downstream effects, such as overdue invoices or additional queries
• User uptake, operating cost and any control failures

Compare the results with the current process. Expand only when the evidence shows that the change saves useful effort, maintains the required controls and is manageable for the people who will operate it.