Your AI strategy is not a ladder

A ten-part series on the AI Value Framework

, by Matt Hartzman

This series walks through the six levers of the AI Value Framework, one post at a time, then turns to buying versus building and how to test each investment.

Your AI strategy is not a ladder

An owner drafts proposals with AI. A manufacturer uses tailored AI to predict production failures.

Which has the better AI strategy?

You cannot tell from those descriptions. You need to know what each business needs—and whether the investment delivers it.

Our AI Value Framework offers six levers:

  • Make people faster.

  • Make people better.

  • Make processes better.

  • Perform bounded work.

  • Create customer value.

  • Create competitive advantage.

The proposal and manufacturing examples sit at different places in that framework. They do not sit on opposite ends of a scoreboard.

A disciplined choice: give salespeople approved tools to prepare account briefs because preparation is consuming time they need for customers.

Not a disciplined choice: commission an agent because ordinary assistants no longer sound ambitious enough.

The six levers can overlap. A better quoting capability could improve a workflow, benefit customers, and become hard to copy. Each claim still needs its own evidence.

The next six posts draw the boundaries using the presentation’s practical examples.

Start with the problem. Nobody gets a prize for choosing the highest number.

AI can make people faster: the proposal still needs an owner

The owner has a prospect to research and a proposal to write.

AI helps with both. The owner checks the facts, adjusts the offer, and sends it.

That is “make people faster”: use AI on demand to remove effort from an individual task.

A mid-sized-business example follows the same logic. Sales teams use approve tools to prepare consistent account briefs and meeting materials. More people, same lever.

Now change the assignment.

An agent checks company systems, identifies accounts needing follow-up, prepares messages, and sends them within approved limits.

That is not merely faster preparation. Work has been delegated. It belongs under “perform work,” with different controls.

And buying licenses without changing how anyone works? That is purchasing, not demonstrated productivity.

For the proposal example, measure the complete task—including fact-checking and corrections. Then ask what the released time makes possible: an earlier response, another proposal, or more time understanding the prospect.

The useful result is not a draft that appeared quickly. It is a good proposal that took less effort to deliver.

AI can make people better: bring better questions to the accountant

An owner receives the monthly financials and knows something looks wrong.

Instead of asking AI to make the report shorter, the owner uses it to explore the results and prepare questions for the accountant.

That is the presentation’s small-business example of “make people better.” Used as an ongoing role partner with reliable business information, AI helps the owner make better use of scarce expertise. It does not replace the accountant.

For a mid-sized business, the example is a finance team using AI to analyze variances, draft commentary, and support business-unit reviews.

The work is not finished when the commentary sounds convincing. Someone still needs to verify the sources and test the explanation.

A near miss: AI reformats the monthly report and finishes sooner, but the owner understands no more than before. That may save time. It does not establish better judgment.

A worse miss: accepting AI’s explanation as the final accounting conclusion.

Ask what changed in the review. Were missing facts surfaced? Were alternative explanations examined? Did the accountant receive sharper questions?

A polished explanation is easy to admire. A better-informed decision is the result worth paying for.

AI can make processes better: stop making the employee the integration

A customer writes to the help desk.

AI classifies the request and drafts a reply inside the existing help-desk software. The employee reviews the response and handles exceptions.

That is the presentation’s small-business example of “make processes better.” AI performs defined steps where the work already happens.

The mid-sized-business example connects customer service, CRM, and order systems so case history can be summarized, requests routed, and resolution supported.

The distinction is the workflow—not how impressive the generated text sounds.

A near miss: an employee copies every ticket into a separate chatbot, asks for a summary, and pastes it back. That may improve individual productivity. It is not the same as embedding AI into the process.

Another miss: accurate summaries that still arrive in the wrong queue. The task improved; the process has not yet earned that claim.

Follow the request from arrival to resolution. Check routing, handoffs, review, exceptions, and fallback procedures. Measure the whole route, not just the summarization step.

If the employee still has to move everything between systems, you have made the employee the integration.

AI can perform work: the reminder is only one part of the job

“Write a polite overdue-invoice reminder” is a drafting request.

“Monitor overdue invoices, review the relevant account information, and prepare reminders for approval” is a delegated assignment.

The second is a small-business example of “perform work.” A bounded agent carries out a multi-step task using company information. A person still approves the message.

Autonomy does not require permission to do everything.

The mid-sized-business example extends the assignment across accounting and CRM: coordinate collections, send approved reminders within defined limits, and escalate exceptions.

What is not this lever? A bookkeeper manually selects an invoice, pastes details into a chatbot, and asks for better wording. Useful drafting assistance, but the bookkeeper is still executing the workflow.

Nor is unrestricted collections authority the objective. Permission to send a routine reminder is not permission to settle a dispute, change payment terms, or write off a balance.

Before deployment, specify system access, permitted actions, approval thresholds, monitoring, escalation, and who can stop the agent.

Do not ask only whether the reminder reads well. Ask whether the agent completed the assigned work—and stayed inside its authority.

AI can create customer value: what changes after the report arrives?

A service firm adds an AI-assisted client report or personalized recommendation to its offering.

That is the small-business example of “create customer value”—provided the addition gives the client something useful, not just more pages.

The mid-sized-business example goes further: an AI-enabled monitoring service or subscription identifies opportunities and risks between engagements.

The customer no longer has to wait for the next scheduled review to receive that service. That changes the offering.

A near miss: the firm uses AI to produce the same report, at the same quality, sooner internally, while the client receives exactly the same service. That is an efficiency opportunity. It does not, by itself, demonstrate additional customer value.

If faster production also means the client receives advice sooner and can act earlier, both levers may apply. Name the customer benefit rather than assuming it.

Test the offer with customers. Check usefulness, demand, willingness to pay, reliability, delivery cost, and support requirements.

The test is not whether the firm used AI. It is whether the customer now gets something worth having.

AI can create competitive advantage: what can’t the next buyer copy?

A specialty business combines its know-how and customer history with configurable AI to improve quoting or service.

A manufacturer combines production and quality data with tailored AI to improve failure prediction and process performance.

These are the presentation’s examples of “create competitive advantage.” In both, the strategic claim depends on distinctive business assets—not simply access to AI.

For the specialty business, ask what a competitor would need to reproduce the quoting capability. The same software may be available. The accumulated knowledge and customer history may not be.

For the manufacturer, test whether its data and operating expertise produce better predictions and better decisions in practice.

A near miss: buy the same forecasting package available to every competitor, use generic inputs, and call it proprietary intelligence. It may improve operations. That does not make it hard to copy.

Another miss: build a custom model that delivers no distinctive business result.

These examples are candidates for advantage, not proof of durability. Verify performance, economics, ownership rights, and how readily the capability can be reproduced.

An improvement earns its operating case. A hard-to-copy improvement must also earn its strategic case.

Buy the software. Build the part that matters.

The specialty business’s quoting example does not start with building a foundation model.

It starts with business know-how, customer history, and configurable AI tools.

That detail matters. Proprietary value can come from what you bring to a product, not from recreating the product itself.

The framework separates four delivery choices:

  • Buy general capability.

  • Configure it for the role or task.

  • Integrate it where the work crosses systems.

  • Build selectively where a distinctive requirement justifies it.

The help-desk example may begin with features already inside the firm’s software. The customer-service example may require connections to CRM and order systems. The manufacturer’s prediction example may warrant tailored capability around production and quality data.

Different needs. Different delivery decisions.

A near miss: building a custom writing assistant to prove that the company takes AI seriously.

Another: buying a tool and calling the job finished when employees still have to move information manually between systems.

Choose the delivery approach after identifying the value and the missing capability. Include maintenance, ownership, and supplier dependency in the decision.

Spend custom-development money on a business difference you can explain.

The collections agent needs more than a content policy

A finance assistant drafts commentary on monthly results.

A collections agent sends a reminder to a customer.

Both produce text. Only one has permission to contact someone on the company’s behalf.

Govern them accordingly.

The finance example needs reliable source information, controlled access, verification, and an accountable reviewer.

The collections example also needs explicit action limits, approval thresholds, monitoring, escalation, and stop or revoke controls.

Then consider the monitoring subscription from the customer-value lever. Now customers depend on a service. Quality, disclosures, data rights, support, and remediation belong in the operating plan.

A near miss: one acceptable-use policy treated as sufficient for all three.

Another: “a human is in the loop,” without naming what that person reviews, when approval is required, or who responds to an incident.

Start with permissions. Can the system advise, update a record, contact a customer, or commit the company?

Then work backward from a plausible failure. Who would notice? Who would intervene? What would need repairing?

A sentence about oversight is not the same as an operating control.

Six levers. Six different investment tests!

You cannot judge a proposal assistant and a manufacturing prediction capability by the same usage chart.

The six examples in this series need different tests:

  • Proposal preparation: less total effort, without losing quality.

  • Monthly finance review: better-supported analysis and better questions, verified by accountable people.

  • Help-desk workflow: better routing, handoffs, and end-to-end resolution.

  • Collections agent: bounded work completed, with exceptions escalated and no unauthorized actions.

  • Client monitoring service: customer demand, useful results, retention, and viable margins.

  • Manufacturing prediction: measurable operating improvement and a capability competitors cannot readily reproduce.

These are proposed tests, not reported results.

A near miss: count licenses, prompts, and pilots, then label the total “business impact.” Those figures describe activity. They do not answer the investment questions above.

For each initiative, name the business owner, establish a baseline, define the risk boundary, and decide what evidence would justify further spending.

An initiative can serve more than one lever. Give each claimed benefit its own test rather than counting the same improvement twice.

Fund the opportunities that fit the business. Scale the ones that earn it. Close the ones that do not.

This is the last post in this series; prior posts:

1) 6 AI Levers that Produce Value for SMBs - https://lnkd.in/gH5VBKX8

2) AI Can Make People Faster - https://lnkd.in/p/gPdQtymd

3) AI Can Make People Better - https://lnkd.in/p/gm288fhk

4) AI Can Make Processes Better - https://lnkd.in/p/g5Rpu35b

5) AI Can Perform Work - https://lnkd.in/p/gtAphWKB

6) AI Can Create Customer Value - https://lnkd.in/p/gaERzPxe

7) AI Can Create Competitive Advantage - https://lnkd.in/p/gfYmwmht

8) AI: Buy the Software, Build What Matters - https://lnkd.in/p/gUq2CfDM

9) AI Governance: More Than a Content Policy - https://lnkd.in/p/gTT9YgZ3

Tell us what’s on your agenda.

A first conversation takes 30 minutes. You’ll leave with a clearer view of the risk, whether or not we work together.