Written by Mostafa Elkourechy, Director of AI & Consulting Operations, SDG Group Middle East
THE DECISION LAYER Series - Article 3
Any enterprise can now get an AI agent to draft a solution in minutes. What separates a good draft from a risky one is the experience built into the agent before it starts, and the expert who checks it after.
Picture two enterprises asked to design the same data platform. Both use the same AI model, and both have a draft in an afternoon. The first enterprise’s agent starts from a blank prompt. The second enterprise’s agent starts from reference architectures, naming and governance standards, and a list of the ways similar projects have gone wrong before. Both drafts look finished.
Which would you sign? The difference is not the model, which both enterprises can rent. The difference is what the agent was standing on when it began.
In the first two articles of this series I argued that agents should draft and humans should release, and that a named person must own the definition behind every number. This article asks the delivery question that follows from both.
What is the agent standing on, and who checks what it builds?
The entrepreneur’s instinct is right: use AI everywhere and ship sooner. The enterprise’s instinct is also right: someone signs the work, and the client holds that someone accountable.
The research supports both. In a preregistered experiment with 758 consultants at Boston Consulting Group, those using GPT-4 on tasks inside the tool’s capabilities completed 12.2% more tasks and finished 25.1% faster, with higher quality. On one complex task outside those capabilities, AI users were 19% less likely to produce a correct solution than those without AI (Dell’Acqua et al., 2023). Same tool, same people. Whether it helped depended on which side of the line the task fell, and the line was not always visible to the user.
A study of 5,179 customer-support agents found an average productivity gain of about 14%, and about 34% for novice and low-skilled workers. The authors describe suggestive evidence that the AI model disseminates the best practices of more able workers (Brynjolfsson, Li and Raymond, 2023). In a 2025 trial, 16 experienced open-source developers took 19% longer with AI on their own repositories, although they expected to be 24% faster (METR, 2025). That sample is small and the tools have moved on, so read it as a caution that felt speed and measured speed can differ.
In my reading, the lesson is the same across all three. AI is strongest when it carries the accumulated know-how of people who have done the work before, and weakest when it is asked to improvise where that know-how is missing.
The model is available to everyone. The experience is not.
Drafting has become a commodity, because every enterprise can reach the same models. What stays scarce is context: the design principles, architecture patterns and hard-won lessons a enterprise has accumulated over years of delivery, in a form an agent can use.
In my view, that context is what keeps an agent’s task inside the zone where it performs well. It tells the agent what good looks like, which patterns apply, and which mistakes to avoid. Experience stops being something that lives in a few people’s heads and becomes part of the delivery mechanism.
1. Embed. The firm's experience is written down so that agents can apply it: design principles, reference architectures, reusable patterns, governance standards, known failure modes and acceptance tests. Each has an owner and a version.
2. Generate. Agents draft the solution inside those patterns. This is the part that has become cheap, so let it run at machine speed.
3. Verify. A named subject-matter expert reviews the draft against the acceptance tests and the client’s context, and signs it off with the evidence recorded. Review depth matches risk, and anything that reaches a client decision gets an expert.
The loop closes when review feeds back. Every error an expert catches becomes a new rule or pattern, so the next draft starts from a higher baseline.
A draft from an agent with no embedded experience, checked by one person, is the weakest version of this design. The reviewer has to rediscover the basics from memory, under time pressure, with thousands of lines to read. Quality then depends on one person’s attention on one day.
With experience built in first, the expert’s time goes where it is most valuable: checking fit to the client’s situation, judging the trade-offs and deciding the exceptions. The patterns catch the known problems. The expert handles the new ones.
Each layer should be visible to you, with a named owner:
|
Layer |
What goes in |
Who is accountable |
|
Embed |
Design principles, reference architectures, reusable patterns, governance standards, acceptance tests |
The enterprise’s practice owners |
|
Generate |
Agent drafts built inside those patterns |
The delivery team |
|
Verify |
Expert review against the acceptance tests and the client’s context, with sign-off recorded |
A named subject-matter expert |
Three questions to ask before you buy AI-assisted delivery:
Fast drafts are becoming a commodity. What will stay scarce is delivery that combines a enterprise’s accumulated experience, applied consistently through agents, with expert review before anything reaches the client. A buyer who asks only for a lower rate will get faster drafts. A buyer who asks for embedded experience and expert sign-off will get a solution they can defend.
In my work at SDG Group, this is the delivery model I help clients adopt: SDG’s delivery experience, design principles and architecture patterns built into agentic delivery itself, followed by review from subject-matter experts. The client receives one end-to-end solution, not a draft and a signature.
If you are rethinking how delivery is designed and signed off, I’d like to compare notes. Reach out and let’s start this conversation.
The two-firm scenarios and the layer table are illustrative examples, not a client case