An AI feature can turn a customer integration from an engineering task into a compliance review because it changes where data goes, how it is processed, and who may see the output. The code may take two days to ship, while approval depends on questions nobody asked before the demo.
Consider Ama, an invented composite of founders I have worked alongside. At 10:14 on Wednesday morning in Accra, she was holding a mug of tea and rereading an email from a customer’s legal team. The subject line contained the product name and “AI assessment required.”
On Tuesday, the customer’s product lead had applauded Ama’s demo. Her feature could read incoming support requests, draft a response, and route urgent cases to a human. The integration had looked straightforward: connect the customer’s help desk, test the workflow, then agree on a launch date.
By Wednesday, the customer wanted to know which model processed the text, where prompts were stored, whether the provider used inputs for training, how long logs remained available, and what happened when a draft contained personal information.
Ama had promised her team that this contract would cover the next hiring decision. Now the integration might miss the customer’s internal budget window. If legal refused the AI workflow, the deal could shrink to a smaller automation project or disappear.
The applause had confirmed that the feature was useful. It had said nothing about whether the customer could approve it.
The feature changed the route to production
Before the AI feature, Ama’s integration moved data between systems according to rules her team could describe line by line. A request arrived. A condition was checked. A ticket moved into the appropriate queue.
The new version sent customer text to a model and returned generated language. That introduced another processor, another set of retention questions, and an output that could vary even when the input looked similar.
From Ama’s side, this was one feature added to an existing workflow. From the customer’s side, the category of the purchase had changed.
That distinction matters when runway is short. A founder may estimate the build from the visible engineering work and treat approval as paperwork that follows. The customer sees a different sequence. Security, privacy, procurement, and legal may need to decide whether the feature can enter production at all.
The more impressive the demo becomes, the more questions it can create. A workflow that summarizes internal messages may expose confidential information to a model provider. A drafting assistant may produce language a staff member sends without checking. An automated recommendation may influence a decision the customer must later explain.
None of those questions proves the feature is unsafe. They prove that usefulness and approvability are separate product requirements.
Compliance questions belong before the demo
Ama’s mistake was not adding AI. She had treated model selection as an engineering choice when it was also a sales constraint.
The customer’s legal email forced her to map the feature more plainly. What data entered the workflow? Which fields did the model need? Could names and account details be removed first? Where did each request travel? What did her team log, and for how long? Which action still required a person?
That map exposed a useful option. The model did not need the full support request to suggest a category and draft a response. Ama’s team could remove identifying fields before sending the text, keep final approval with an employee, and disable model input retention where the provider supported it.
This was the turn, but it arrived late. The customer had already moved the integration out of the product lead’s hands.
I have seen the same timing problem in other product decisions. A working demo feels like evidence that the hard part is over. Yet, as with the demo that worked while Monday’s invoice threatened runway, technical success can expose the commercial constraint that was hidden before anyone saw the product.
Build an approval path beside the feature
Ama did not respond with a broad claim that the system was secure. She sent a short data-flow diagram, named the model provider, described what was retained, separated automated actions from human approvals, and listed the configuration changes her team could make.
She also divided the proposed integration into two stages. The first used rules-based routing with no generated response. The second added the drafting feature after the customer completed its review.
That reduced the size of the decision. The customer could continue the integration without approving every AI use at once. Ama could preserve part of the contract while legal examined the feature that had changed the risk.
This is the practical work to do before the next demo. Write down every system that receives customer data. Mark what leaves your infrastructure. Identify which output could affect a person, payment, record, or customer communication. Then decide which parts can launch without the model.
You may discover that the AI feature needs less data than your prototype sends. You may find that the customer requires a human approval step. You may also learn that the feature cannot fit the buyer’s current policies. Finding that before applause is cheaper than discovering it after a launch date has entered the forecast.
The second demo answered a different question
By the time Ama returned to the customer call, she had changed what she showed.
The first demo had answered, “Can this work?” The second answered, “Can your company adopt it without accepting an undefined risk?”
She showed the redacted input, the employee approval screen, and the point where the rules-based workflow could operate while the AI review continued. The product lead still liked the generated draft, but the most important moment came when the legal contact stopped asking where the data went. The diagram already showed it.
The deal was still unresolved. Legal had not approved the AI feature, and Ama did not pretend otherwise. But the original integration could move forward, the customer knew exactly what remained under review, and her team no longer had to defend a black box with a polished demo.
On Thursday morning, Ama replaced one line in the next sales deck. Under the AI feature, she added four words: “Data path available here.”
Comments
No comments yet.