How to Evaluate AI Tools

A useful AI tool needs to hold up through revisions, work within clear permissions and produce results a team can confidently review.

That’s the practical thread connecting several recent releases from OpenAI, Google, Adobe and Thinking Machines. Creative tools are adding editing controls. Enterprise assistants are making business information accessible through familiar interfaces. New models are creating additional customization options.

For founders and marketing teams, these developments make it worth asking more precise questions before adopting another tool.

Can it preserve approved details? What information can it access? Can it recommend changes, execute them, or both? What does the complete workflow cost?

Creative AI needs to survive the next edit

OpenAI’s current developer documentation describes GPT Image 2.6 Sunburst and Flare, with complexity controls, markup-guided edits, SVG output and animated GIF output through the Image API. Complexity is separate from rendering quality. Higher complexity can increase latency and introduce unwanted changes, so compare settings on your own task. OpenAI image generation guide (opens in a new tab)

For a marketing team, the useful evaluation is specific: when someone requests a different background, does the approved product, lettering and layout remain intact?

Unexpected changes create additional review work. That effort belongs in the evaluation alongside generation speed.

Video brings the same challenge into motion

Google’s November 5 Gemini Omni 1.2 Flash release adds clips up to 15 seconds and conversational editing. API and business availability are general, while Gemini-app access is rolling out to Google AI Ultra subscribers. Google’s claims about improved reference consistency still require testing against a business’s own assets. Google announcement, November 5, 2026 (opens in a new tab)

Short explainers, campaign concepts and product clips are possible evaluation areas. Teams should assess the full path to an acceptable video: generation, revisions, review and final editing.

Generation fees are only one part of that cost. Revisions, review time and final editing belong in the total.

An integration’s permissions define its practical value

Adobe announced expanded CX Enterprise Coworker integrations on November 4, including ChatGPT, Claude, Microsoft Copilot, Databricks and Google’s business AI platform. Adobe announcement, November 4, 2026 (opens in a new tab)

Its implementation documentation provides important qualifications. The external-platform experience returns read-only insights and requires provisioned Coworker access, Adobe Experience Platform and a qualifying associated application. Usage consumes existing AI credits, with no separate Adobe integration charge. Adobe Experience League: Coworker in AI platforms (opens in a new tab)

This illustrates a distinction every business should make when evaluating an assistant:

• Read: retrieve information from an authorized system.
• Recommend: suggest a decision or change.
• Execute: modify data or take action.

Confirm which capabilities an integration actually supports and whose permissions govern its access. A conversational interface does not establish authority to change a campaign or customer record.

We walk through these levels in more detail, with a checklist, in our earlier guide.Read: Can your AI integration read, recommend or act?

Model availability has several meanings

Thinking Machines released Inkling 2 Small on November 5. It accepts text and images, supports adjustable reasoning effort and provides inference and fine-tuning through Tinker. Thinking Machines: Inkling 2 Small (opens in a new tab)

The company plans downloadable Apache-2.0 weights after external safety testing, targeting the following week. At the research cutoff, the weights remained closed.

Hosted access, fine-tuning access and downloadable weights are separate milestones. Build project plans around the capability available today, and track promised availability separately.

A practical pilot: request one change

Untested suggestion: use an approved image-editing tool your team already has access to.

1. Choose a product image or campaign graphic you have permission to use.
2. List five details that must remain unchanged: shape, logo, wording, colors and placement.
3. Request one edit, such as replacing the background.
4. Compare the result with the original against that checklist.
5. Repeat twice, recording unwanted changes, manual corrections, elapsed time and charges.

A starting prompt:

“Replace only the background with [description]. Preserve the product’s shape, proportions, lettering, colors and position. Leave every other element unchanged.”

The prompt expresses your requirements; the review establishes whether the tool met them.

Our recommendation is simple: define “acceptable” before generating. A small, documented pilot can reveal where a tool fits, where it needs supervision and whether it deserves a larger role in your business.

What must remain consistent in your team’s AI workflow? Start your next evaluation there.

If you want help designing that kind of evaluation for your own team, take a look at Build With AI.Explore Build With AI