Week of 14 September 2026
Which parts of your AI system will gain value, and which may no longer be needed?
Build company knowledge that becomes more useful as models improve.
DeepSeek reaches a similar design-evaluation result at about 70 times lower estimated model cost
OpenDesign tested models on prototype design tasks. DeepSeek V4.1 Flash scored close to GPT-6 Astra while costing a small fraction as much to run on the same work.
Design-evaluation score
Out of 100. Higher is better.
- DeepSeek V4.1 Flash
- 81.2
- GPT-6 Astra
- 82.7
Estimated model cost per design
US dollars. Lower is cheaper.
- DeepSeek V4.1 Flash
- $0.023
- GPT-6 Astra
- $1.61
Source figures dated 12 September 2026. Scores cover this prototype task set. Costs are estimated from model usage and exclude human review and implementation.
Test the work that exists only because today's model needs help
Some stages compensate for current model limits. Keep them while they earn their cost back, and test them again when the model changes. DeepSeek's own model card reports a 19.8 percentage point gain on a software-issue benchmark between V4 Flash and V4.1 Flash, at the vendor's chosen settings, so the gap a workaround covers can close inside one release.
- Stage that compensates for a model limit
- Stage that stays in both paths
Older workflow
- Extract sections
- Summarise separately
- Compare
- Human approval
A worked example of one document-comparison workflow, not a universal process. The first three stages exist chiefly because an earlier model compared long documents poorly.
Test with a stronger model
- Compare directly
- Human approval
A stronger model may make extraction and separate summaries unnecessary. Establish that on your own representative documents before removing a stage.
Foundations that remain
- Access rules
- Verified source material
- Decision authority
Better reasoning does not grant permission to read a customer record, and it does not grant authority to make a commercial commitment. These stay in every version of the workflow.
The point is to test whether model-specific steps still earn their keep. It is not an instruction to remove controls. An investment case that assumed years of continuing advantage from a workaround needs revision once the model makes that step unnecessary.
A model release reaches real work only after it crosses your adoption route
A public release does not place a model inside your approved systems. The assistant your staff use sits on top of a model you can swap, and the distance between the two is where the improvement waits.
Public model release
The vendor publishes the model and the benchmark figures. Nothing in your company has changed yet.
The lag between release and real business use
Approved supplier availability
Can the company obtain the model through its approved platform?
Data, security and risk assessment
Can the intended information and use case be handled within company and regulatory boundaries?
In Hong Kong financial services, the Monetary Authority's generative-AI research paper recommends red-teaming and penetration testing inside technology risk management. The Securities and Futures Commission's November 2024 circular binds a narrower group: licensed corporations using language models in regulated activities, which validate before approval and on material change. Confirm current obligations with legal and compliance.
Representative workflow test
Does the complete application work with the existing systems, permissions and failure controls?
Business owner and staff readiness
Who approves consequential actions, challenges errors and resolves issues?
Use in real work
Staff apply the model to work that matters, with a named approver and a way to raise a problem.
What shortens the next crossing
Record the systems the assistant reaches, the approvers, the failure cases and the test documents once. The next model then needs an assessment of what changed rather than a fresh route from the beginning.
What turns an AI suggestion into reusable company knowledge?
A useful suggestion needs more than a capable model. It needs verified context, a responsible business decision, an action in the operating system, and feedback about what happened afterwards.
What the company learns at the last stage becomes the context the assistant detects with next time.
Illustrative case: a hotel listing correction
One worked application of the same four stages. The assistant suggests and records while the business keeps authority.
- Detect
A guest complaint about a room feature conflicts with the verified room specification.
- Decide
Marketing and operations determine whether a room-category exception applies.
- Act
The responsible owner amends the booking-platform listing, or declines the change with a recorded reason.
- Learn
The company checks later complaints and relevant booking context before treating the correction as a reusable lesson.
Whether the guest stops complaining tells you much less than whether the same complaint comes back.
What should remain useful when the model changes?
Six company-owned assets carry the value of past work across a model or supplier change. None of them belongs to the vendor.
These records can be corrected, governed and made available to the next approved model.
In the hotel illustration
The retained record includes the verified feature, the affected room categories, the approved or rejected listing correction, the reason for the decision, and the later evidence that may confirm or challenge the explanation.
Keep the employee response separate from the explanation and the outcome. Review samples of accepted suggestions as well as disputed ones, and retire a conclusion when new evidence overtakes it.
Three questions for the next AI proposal
Each one returns to a link in the same chain: what a better model may displace, what stands between a release and real use, and what the company keeps either way.
Which parts of this system exist mainly because the current model cannot yet do the work directly?
Fund those stages against a plausible payback period, then retest them as models improve.
What would it take for us to test and adopt a better model safely?
Identify supplier availability, data and security boundaries, representative business tests, responsible owners and workflow changes.
Which facts, decisions and feedback should remain useful if we change the model or supplier?
Preserve these in usable company records rather than leaving them in isolated chats or project notes.