Ask ten people in tech to name the top AI companies and you’ll get the same six names: OpenAI, Google DeepMind, Anthropic, Microsoft, Nvidia, Meta. Ask which of those they personally pay money to, and the room goes quiet. Recognition is easy. Procurement is not.
That gap is what this playbook closes. Below is the sequence I’ve watched work at a 12-person bookkeeping firm, a 60-person garden-tool retailer, and a 400-person logistics operator. No vendor deserves to be chosen before you’ve finished step one.
Step 1: Write the job down before you write the vendor down
Every failed AI purchase I’ve seen started with the technology. Every good one started with a sentence containing a person, a verb, and a number.
Weak version: “We want to use AI.” Strong version: “We want Priya’s four-person support team to resolve 45% of inbound product questions with no human, in under four seconds.”
That second sentence makes the decisions for you. It sets a latency budget. It sets an accuracy bar. It names an owner. It also tells you whether you need a general chatbot, a voice agent, or a forecasting model, because those are three genuinely different purchases wearing the same acronym.
Sort the work into three buckets
- Conversation: support chat, phone lines, email triage, booking.
- Documents and knowledge: contract summarising, policy search, first-draft writing.
- Prediction and numbers: demand forecasting, lead scoring, anomaly detection.
Each bucket pulls you toward a different shortlist. Conversation work points at OpenAI, Anthropic, or Google. Document work often lands on Microsoft’s stack because the files already live in SharePoint or OneDrive and nobody wants a second copy of the employee handbook floating around. Forecasting has almost nothing in common with chatbots, and that’s where a platform such as DataRobot and its AutoML approach to turning existing data into decisions starts to make more sense than another chat subscription.
Step 2: Shortlist two vendors per job, never ten
Sales teams will happily send you twelve decks. You should read two. The filter that matters most is boring: does the thing have a documented API, a published security posture, and a support channel that answers on a Saturday? If a vendor can’t produce a SOC 2 report and a rate-limit page, they’re out before the demo.
If you want a current map of who is actually shipping product versus who is mostly publishing announcements, a 2025 rundown of which AI companies are leading and which are losing ground is a reasonable starting filter. Read it as a shortlist, not a shopping list. A company with the best benchmark scores can still be the wrong pick if their pricing model punishes you at volume.
Step 3: Run a two-week pilot behind a hard threshold
The garden-tool retailer ran theirs on live traffic, not demo questions. Last quarter’s 400 weekly support tickets went through a chatbot, and the team wrote down three numbers before the first test: 45% deflection, under 90 seconds median resolution, CSAT no lower than 4.2 out of 5. If the pilot missed all three, the vendor was out and they moved to the second shortlist entry.
Two things kill pilots. Cherry-picking twenty easy questions, and running for a month with no numeric target, so everyone argues about vibes at the end. Load real tickets. Set the number first. A workable five-step workflow for using a chatbot in real operations covers the setup properly if you want the checklists.
Step 4: Deal with the conversations that aren’t typed
Voice is where the leading AI companies diverge hardest, and where most buying guides stay vague. The constraint isn’t intelligence, it’s latency. Anything past roughly 700 milliseconds of dead air on a call and people assume the line dropped. They hang up, and you’ve spent money to lose a customer.
The retailer added a phone line for order status in month two. Roughly 1,800 calls a month, mostly “where is my delivery.” Their build followed a six-week plan for launching a Poly AI voice agent, which front-loads the boring parts, telephony and fallback rules, before any prompt engineering happens. Skip that ordering and you’ll have a clever agent that can’t transfer to a human.
Step 5: Fix the plumbing before you scale anything
This is the step enterprises regret skipping. If your data can’t leave a specific region, or your legal team wants a signed data processing agreement, your shortlist collapses to whatever runs inside infrastructure you already trust. For a lot of organisations that means going through Microsoft rather than calling a model provider directly.
It’s worth understanding what Azure OpenAI Service includes, how its token pricing and provisioned throughput work, and when it beats going direct before you commit, because provisioned capacity is a commitment, not a pay-as-you-go tap. A bank I worked with saved about 30% by switching two high-volume workloads to provisioned throughput and leaving the spiky one on pay-as-you-go. Same model. Different purchasing decision.
Step 6: Let someone without an engineering degree ship something
Not every purchase needs a pilot. Campaign pages, seasonal landing pages, and small-business sites are now fast enough to be a marketing task rather than a ticket in a developer backlog. A three-person brewery near me published a page for a limited seasonal release in an afternoon by following a 90-minute walkthrough for building a small business site with an AI website builder. Nobody wrote code. The page converted at roughly the same rate as one their agency produced for eleven times the cost.
The test for this category is simple: if a competent marketing generalist can’t get a decent result inside two hours, the tool is either oversold or wrong for the job.
Step 7: Put a review date on every tool you buy
Model updates land without warning. Pricing changes. A vendor you chose in March can quietly become the expensive option by September. Book a quarterly review with three questions and answer them honestly:
- Did the metric we set in the pilot hold, improve, or drift?
- What did we pay per useful outcome, not per seat?
- Would we buy this again at today’s price and today’s capability?
Answering question two is the one most teams avoid, because a $20 monthly seat feels free while an API bill of $2,400 does not. Divide the bill by the number of tickets deflected, calls handled, or hours returned. That number is the only honest comparison between vendors.
What the assembled stack looked like
Eighteen months in, the garden-tool retailer runs four things. Chat on a frontier model for product questions. A voice agent for order status. An AI website builder for campaign pages. A forecasting model trained on their own two years of sales history to plan stock for spring.
Deflection sat at 38% in month one and 61% by month three, mostly because they rewrote the top twenty answers after reading real transcripts. Total spend landed around $2,400 a month. The team stopped counting hours saved once it passed roughly one and a half full-time equivalents, which is when a purchase stops being an experiment and starts being headcount planning.
The habit worth copying isn’t any single vendor on that list. It’s keeping a changelog for every AI tool you run, noting the date you last changed a prompt, retrained a model, or renegotiated a rate. When something breaks in month nine, that file is the difference between a two-hour fix and a two-week mystery.