Why Most SMB AI Strategies Are Really Shopping Lists

By Victor Hoang · Updated Oct 10, 2026

Ask a leadership team at a 60-person company for their AI strategy and you'll usually get an inventory. A chat assistant for everyone. A meeting note-taker. An AI feature switched on in the CRM. Maybe a line about "exploring agents in Q1." It sounds like progress, and it is a purchasing plan.

A list of tools is not a strategy. A real AI strategy names the few workflows and decisions the company will change, who owns each one, and what result will prove it worked. Everything else is shopping. This piece gives you a way to tell the two apart, a test you can run on your own plan in an afternoon, and a short sequence for turning the list into something an executive team can actually be held to.

Key Facts

• Strategy scholar Richard Rumelt's "kernel" has three parts: a diagnosis, a guiding policy, and coherent actions that carry the policy out. A tool list supplies none of the three on its own.
• In MIT's NANDA report on enterprise generative AI, as covered by Fortune in August 2025, 95% of companies in the dataset were falling short on implementation. The report's authors blamed a "learning gap" in how organizations integrate tools, not model quality. It was a preliminary study of a limited sample.
• The same coverage reported that purchased tools and partnerships succeeded about 67% of the time, while internal builds succeeded about one-third as often, and that the report's lead author credited successful startups with picking one pain point and executing well.
• In a study of 5,179 customer support agents, an AI assistant raised issues resolved per hour by 14% on average and by 34% for novice and low-skilled workers, with minimal impact on experienced workers. The gain was measured in one named workflow, not across "the business."

Why the shopping list feels like a strategy

Shopping lists are comfortable for three reasons.

First, buying a tool is a visible, finishable action. You can announce it, budget it, and tick it off. Choosing what not to do is invisible and politically harder.

Second, tools arrive faster than decisions. A new assistant ships every month, and each one is easy to justify on its own. A list grows by accretion, while a strategy has to be chosen.

Third, nobody is accountable for a tool being "adopted." Adoption is a vague noun. If the plan says "roll out an AI notetaker," the plan succeeds the day the licenses are assigned, whether or not a single meeting gets better.

Rumelt's critique of bad strategy applies directly. In a summary of his book, a vision or a goal on its own is not a strategy, because without concrete steps the organization is left rudderless. "We will adopt AI across the business" is a vision. A tool list is the shopping trip taken in its name.

The kernel, applied to AI

Rumelt's three parts translate cleanly. Here they are in AI terms, with the question each one forces.

Diagnosis: what is actually hard or expensive for us? Not "we're behind on AI." Something specific: quotes take three days, so we lose deals to faster rivals. Support tickets bounce between two people because nobody can find the last customer decision. Month-end close eats the finance lead for a week. A diagnosis points at a workflow with a cost.

Guiding policy: what is our approach, and what are we ruling out? For example: "We will apply AI to repetitive, reviewable work with a clear owner, and we won't automate anything that reaches a customer without a person approving it." The policy is as much about exclusion as selection. It tells the whole company what to say no to, including a team that has found a shiny tool of its own.

Coherent actions: what, specifically, changes, by whom, by when? A named workflow gets redesigned. A named person owns it. Data access, review steps and training are arranged so they support each other instead of pulling apart.

If you can't write the diagnosis in a sentence a skeptical board member would accept, you haven't got a strategy yet. You have a catalog.

Shopping list versus strategy: the test

Run your current plan through this table. Each row is a place where the two diverge in practice.

Score yourself honestly. If most of your answers sit in the left column, that is the common case, and it is fixable.

What the evidence says about where value shows up

Be careful with any single study, but the pattern across the better-documented ones is consistent: value shows up where AI is attached to a specific job, not spread across the organization.

The customer support study mentioned above is the clearest example. The researchers measured a generative AI assistant rolled out in one function, and found a 14% average productivity gain with much larger gains for newer workers and little for veterans. That is a result you can only get by naming a workflow, a metric and a population. A company with "AI for everyone" couldn't produce it, because there'd be nothing to measure it against.

The MIT NANDA findings, which Fortune summarized, point the same way from the failure side. They attribute the shortfall to integration and learning, noting that generic tools are flexible for individuals but "don't learn from or adapt to workflows." They also reported that more than half of generative AI budgets went to sales and marketing tools while the biggest returns appeared in back-office automation. Treat that last point as a prompt to check where your own money is going, not as a rule. The study was preliminary, and its sample was not built to represent small businesses.

What an SMB can take from this: narrow beats broad, and a bought tool wrapped around a redesigned workflow beats an unowned experiment. Both are strategy choices, not tool choices.

A worked example (illustrative)

Take a made-up 45-person wholesale distributor. Its first AI plan reads: "Roll out a chat assistant to all staff, add an AI note-taker to sales calls, trial an inventory forecasting add-on, and explore an order-entry agent."

That's four tools. Now the same company writes a strategy.

• Diagnosis: Quote turnaround averages three business days and customers go to faster competitors. The delay is mostly staff retyping specs from emails and looking up past pricing.
• Guiding policy: Use AI to prepare quotes from past orders and inbound requests; a salesperson reviews and sends every quote. No AI-sent customer messages. Everything else waits.
• Coherent actions: The sales operations lead owns quoting. Past quotes get cleaned into a usable reference set. The review step becomes a defined part of the process. Two reps pilot for six weeks.
• Proof: Median quote turnaround against the current baseline, plus quote win rate and review corrections per quote. If turnaround doesn't drop within the pilot window, the team changes the workflow or stops.

Same company, one bet instead of four, and a result somebody has to report. The other three tools aren't banned forever. They're sequenced behind a result.

Where governance and strategy meet

A strategy that names workflows also makes governance easier, because you know where AI touches decisions. The governance gap is mostly a visibility problem, and a workflow-level plan gives you the visibility for free: each named workflow gets a data tier, a review rule and an owner.

It also connects to the question of what happens when tools change under you. A plan built on workflows survives vendor shifts far better than one built on products, a point we make in an AI strategy that survives model churn. And it makes the ROI conversation tractable, since measuring AI ROI beyond time saved requires a baseline for a specific process, which a shopping list never defines.

The shape of the work matters too. Whether a workflow suits an assistant that suggests or an agent that acts changes the review design, which is the subject of AI copilots versus agents. And if your bottleneck is handoffs between people rather than any single task, coordination cost is the better diagnosis to start from.

What to do next

You don't need a consultant or a quarter. You need a few honest working sessions.

1. Write the inventory. List every AI tool and feature anyone in the company pays for or uses, including the ones inside software you already own. Expect surprises.
2. Run the table above. Mark each row of your current plan as shopping list or strategy. Be blunt.
3. Pick three candidate workflows. Choose ones with a clear cost, a repeatable shape and work that someone can review. Rank them by cost and by how hard the data is to get at.
4. Choose one. Resist the second. Name the owner, who must be a business leader and not just an IT contact. If you're torn about timing, moving now versus waiting for AI to mature lays out how to decide.
5. Write the kernel on one page. Diagnosis, guiding policy including what you rule out, coherent actions, and the proof metric with its baseline and a stop date.
6. Set the review cadence. A short monthly check where the owner reports the metric. Anything not tied to the chosen workflows goes into a "later" list, not a purchase order.
7. Convert the lessons into objectives. Once one workflow has a result, turn the next into a strategic objective and carry it through strategy execution the way you would any other priority.

The point isn't to buy fewer tools. It's that every tool you keep is attached to a job, an owner and a number.

Learn More

• An AI Strategy That Survives Model Churn
• Measuring AI ROI Beyond 'Time Saved'
• The Governance Gap: What Leaders Get Wrong About AI at Work
• Strategy Execution