The 2026 Procurement Playbook

Why procurement changed

Procurement at a growing company used to be an email chain, a shared spreadsheet, and one patient person in finance who knew which vendor had a contract on file. That worked at 60 people. At 300 people it breaks, and it breaks quietly: a team signs a $48,000 data tool on a credit card, a renewal auto-renews at a 9% uplift nobody saw, and a vendor with production access never finishes its security questionnaire.

The usual fix was to buy a procurement suite and staff a procurement team. Most companies under 1,000 people never did either. They tolerated the mess because the work was too small to justify a hire and too scattered to put in one tool. Agents change that math. The work of procurement is mostly reading, chasing, comparing, and routing, and that is now work an agent does well for a few hundred dollars a month.

This playbook comes from five operators who rebuilt procurement this way over the last 18 months. Rosalind Achebe at Tallowmere wrote it with us. Between them they run about $61 million a year in third-party spend across 1,720 employees, with a combined procurement headcount of three. What follows is how they split the work, what they automated first, and which calls they refuse to hand to an agent.

The five jobs

Start by splitting procurement into five jobs, because each one has a different failure and a different place where a person must decide. When teams say “procurement is slow,” they usually mean one of these five is stuck, and the cure for one does nothing for the others. Name them, give each one an owner, and give each one an agent.

1. 01Intake. Someone wants to buy something. Capture what, why, how much, for how long, and who owns it, in the place where they already ask.
2. 02Approval. Route the request to the people who must say yes, in the right order, with the context they need to say yes quickly.
3. 03Onboarding and security review. Collect the vendor’s tax forms, banking details, insurance, and security documents, and decide whether the risk is acceptable.
4. 04Renewals. Know every contract’s end date, notice period, and price, and start the decision early enough to have a choice.
5. 05Payment. Match each invoice to an approved purchase, pay it on time, and stop the ones that do not match.

For each job, ask two questions. First, what does the agent prepare? Second, what does a person decide? An agent that prepares and a person who decides is the pattern that held across every company we talked to. An agent that decides alone works only where the rule is written down and the cost of a wrong answer is small.

Those numbers are the pooled results from our five contributors in their first full year. The approval time matters most to the rest of the company, because it is the number employees feel. The renewal savings pay for the program several times over. The last number is the one your auditor and your board care about.

Intake and approvals

Put intake where people already ask, because a request form nobody opens is worse than no form at all. At Tallowmere, 71% of purchase requests used to arrive as a Slack message to Rosalind. Today they still arrive in Slack, in a #buy channel. The difference is that an agent answers first, asks the missing questions, and turns the thread into a structured request.

The agent asks for five things: the vendor, the business reason, the expected annual cost, the term, and the budget owner. It looks up the vendor in the ERP and the contract folder to see if the company already pays for it. Roughly one request in seven at Tallowmere turns out to be a tool another team already has, and the agent says so before anyone spends a minute on approval.

Approval routing is where most companies over-engineer. Write the policy as a short table: who approves at which amount, and which categories add a reviewer. Below $2,500 a year with an existing budget line, the manager approves. From $2,500 to $25,000, finance joins. Above $25,000, or for any vendor that touches customer data, security and the department head join. The agent applies the table, attaches the context, and chases.

Onboarding and security review

Vendor onboarding is a document chase, and that is exactly why it stalls. Somebody must collect a W-9 or W-8, banking details, a certificate of insurance, and, for anything that touches data, a SOC 2 report or a completed security questionnaire. Each item lives with a different person at the vendor. A human gives up after the second follow-up. An agent does not.

Halvorsen Bio buys from 340 active vendors, from reagent suppliers to cloud software, and Anjali Moorcroft’s team of two reviews every one that gets system access. Their agent sends the request, tracks what came back, reads the SOC 2 report, and writes a one-page summary: the scope, the period, the exceptions the auditor noted, and the sub-processors. It flags anything that does not match the questionnaire answers.

Banking details deserve a separate rule, because they are the most common route for payment fraud. Never let an agent accept a bank account change from an email alone. At every company in this playbook, a change to vendor banking details triggers a call-back to a phone number already on file, made by a person. The agent opens the task and blocks payment until the call is logged.

Renewals on a calendar

Renewals are where the money is, and most companies find out about them from the invoice. A typical software contract auto-renews with a 30-, 60-, or 90-day notice window and a price uplift of 5% to 10%. Miss the window and you have no leverage. The fix is simple to describe and tedious to do: know every end date, and start each decision before the window closes.

Have an agent read every signed contract in your folder and extract the term, the end date, the notice period, the auto-renewal clause, the uplift cap, and the price per seat or unit. Then have it pull actual usage where it can: seats assigned against seats used from the admin console or SSO logs. At Lumenrow, the first pass over 212 contracts found 41 with a notice window inside the next 120 days, and 9 they had no record of at all.

1. 01At 120 days before the notice deadline, the agent posts a renewal brief to the budget owner: cost, usage, overlap, and the uplift in the contract.
2. 02At 90 days, the owner chooses one of three paths: renew as is, renegotiate, or cancel. The agent records the choice.
3. 03For a renegotiation, the agent prepares the pricing benchmarks and a draft position. A person runs the conversation.
4. 04At 30 days, the agent checks that the chosen path is done, and escalates to finance if the contract is about to renew by default.

Do not let the agent send cancellation notices on its own. A notice is a commitment with a counterparty, it is often hard to reverse, and the vendor’s account team will call the person who owns the budget anyway. The agent drafts the notice. The owner sends it.

Paying vendors

Payment is the last gate, and it only works if the earlier jobs did theirs. An invoice should match an approved request, a vendor that passed review, and a price that agrees with the contract. When all three line up, the agent codes the bill, attaches the approval, and queues it for the next payment run. When one does not, the bill stops and a person sees why.

Fernwick processes about 1,900 vendor bills a month. Their agent auto-matches 83% of them to an approved purchase and a contract price, inside a 2% tolerance. The other 17% go to an exception queue with the reason written at the top: no approved request, a price above the contract, a duplicate invoice number, or banking details that changed since the last payment. Imani Castellanos’s team clears that queue daily in about 40 minutes.

• Pay through one system. Cards for subscriptions, bill pay for invoices, and both feed the same ledger.
• Issue a virtual card for each software vendor, with a limit equal to the approved amount. An overage then fails at the card, not in a review three weeks later.
• Hold any bill whose vendor has an open review item. The agent releases it when the item closes.
• Never let the agent approve a payment above its own matching rule. It can prepare the batch; a person releases it.

What stays human

Keep three decisions with people, no matter how good the agent gets. The first is any new vendor commitment above your threshold. The second is negotiation. The third is an exception to policy. Each one either creates an obligation with another company or changes what your rules mean, and the person who makes it must answer for it later.

New commitments stay human because the agent cannot know your strategy. It can tell you that two teams want overlapping analytics tools and that one costs 40% less. It cannot tell you that the expensive one is the vendor your largest customer uses, or that you plan to build the feature yourselves next quarter. Signing is a strategic act. Let the agent make the decision easy, not make it.

Negotiation stays human because it is a relationship. Agents are excellent at preparation: benchmark prices, usage data, the history of every past concession, the date the vendor’s fiscal quarter ends. Our contributors let agents draft the first email and the fallback positions. None of them let an agent sit across the table, and two of them tried it once and stopped after a vendor escalated to their CEO.

Exceptions stay human because each one sets a precedent. When a team wants to skip security review for a two-week pilot, someone must weigh that risk and write down why. If the agent grants exceptions, your policy erodes one reasonable request at a time. Have the agent log every exception, with who granted it and why, and review the log each quarter.

The tools we ranked

Pick tools for the record, not for the feature list. The best procurement stack is the one where an agent can read the request, the approval, the contract, the card, and the bill without stitching five exports together. We asked each contributor to score the tools they have run in production.

Two notes on the ranking. First, no contributor runs all five, and none needs to. A common stack is Ramp or Brex for spend, NetSuite for the ledger, and an agent in between that reads both. Second, the tool matters less than the data inside it. A clean vendor master with one record per vendor beat every feature on this list.

• RampSpend management, corporate cards, and bill pay
• BrexCorporate cards, banking, and spend management

So what do you do?

Start with renewals, because they return money in the first month and need no change in employee behavior. Then fix intake, then payments, then onboarding. Our contributors who started with an intake form spent a quarter on adoption before they saw a dollar. Those who started with the contract folder paid for the whole program before intake went live.

1. 01This week, put every signed vendor contract in one folder and have an agent extract end dates, notice periods, uplifts, and prices into one table.
2. 02This month, post a renewal brief to each budget owner for every contract with a notice window in the next 120 days.
3. 03Write your approval policy as one table of amounts, categories, and approvers. Keep it to one page.
4. 04Open a single intake channel and let an agent turn each request into a structured record and check it against what you already pay for.
5. 05Turn on invoice matching against approved requests, with a written exception reason on every bill that does not match.
6. 06Require a call-back to a known number for every banking change, done by a person.
7. 07Each quarter, review the exception log and the ten slowest approvals, and change the policy or the prompt.

Measure three numbers from the start: request-to-approval time, the share of spend with a contract on file, and the renewal dollars you decided on before the notice window closed. If all three move in the right direction for two quarters, your program works. If they do not, look for the job where a person stopped deciding and started chasing, and give the chase back to the agent.