Can AI Application Development Fail Gracefully?
A bounded AI feature inside a running system can be migrated without downtime if you define its failure mode first. Here is how to judge accuracy, cost and rollback before you commission it.
AI application development that fails gracefully means the feature has a defined fallback when the model is wrong, slow or too expensive. Judge a proposal on three things: a stated accuracy threshold tied to a business action, a per-request cost ceiling, and a rollback path that does not stop operations. Without those, you are buying a demo, not a production feature.
Where does AI actually help a small business?
AI helps a small business at the step where a person reads unstructured input and types a structured result into a system. The value comes from removing that transcription step, not from replacing the judgement around it.
In back-office operations, the pattern repeats: an invoice arrives as a PDF, a supplier email confirms a delivery date, a customer form contains a free-text complaint. AI document processing can extract the fields and propose a record. A person still approves the record.
For a business in Amsterdam, the regulatory detail matters. If the feature touches personal data, the GDPR purpose-limitation principle means the extracted fields must be limited to what the original process needed. if a field is not used by a downstream step, it should not be extracted at all.
The same logic applies to customer-facing features. A reply drafted from a knowledge base is useful when an agent edits and sends it. It is not useful when it sends itself and the customer receives a confident wrong answer about a return window.
If you want to see how these bounded features are scoped, our AI application development page describes the delivery pattern.
How do you integrate an LLM into an existing system without downtime?
You integrate an LLM into an existing system by running it beside the current process, not inside it, until the new path has proven itself on real traffic. The migration is a parallel run, not a switch-over.
The mechanism is a shadow mode. The existing workflow stays the system of record. The AI feature receives the same input, produces a proposed output, and that output is logged but not used. Staff compare the two for a defined period.
- Add an event or queue that copies the input to the AI service without altering the current step.
- Log the AI output alongside the human output for the same input.
- Define the agreement threshold for promotion, for example 95% field-level match on a sample.
- Promote the AI path to a suggestion that a person accepts, then to an automatic path only for the narrow cases that passed.
- Keep the old path available as a switch that an operator can flip without a deployment.
This is where AI integration services differ from a standalone tool. The integration work is mostly plumbing: an API call, a timeout, a retry policy and a queue that absorbs failures. if the AI call fails, the existing manual step must still be reachable in one click.
The delivery team controls the integration code, the logging and the fallback switch. The client controls the acceptance threshold and who is authorised to promote the feature. Write both into the scope before work starts.
How do you know if an AI feature is accurate enough?
An AI feature is accurate enough when its error rate is measured against the cost of a wrong answer in the specific step it serves. Accuracy is not a single number; it is a threshold attached to an action.
For AI document processing, measure field-level accuracy, not document-level pass or fail. A purchase order with twelve fields where eleven are correct is not 92% useful; it depends on which field is wrong. A wrong total is a finance problem. A wrong internal note is not.
Build a test set from real historical inputs before development starts. Two hundred representative documents or messages, labelled by the person who currently does the work, give you a baseline the vendor cannot tune around.
| Decision | What to measure | Threshold that changes the design |
|---|---|---|
| Auto-approve a record | Field-level accuracy on money and identifiers | Below the agreed rate, route to human review |
| Draft a customer reply | Factual error rate on a labelled sample | Any error about price or policy requires human send |
| Extract for internal notes | Document-level completion rate | Lower bar acceptable because a person reads it |
| Trigger a downstream action | False-positive rate on the trigger | High cost of false positives forces confirmation step |
if a wrong answer would cost money, the citation column decides the purchase. A feature that shows the source sentence beside its extraction lets a reviewer catch errors in seconds. A feature that shows only the answer moves the cost to the reviewer.
Accuracy also drifts. New supplier formats, seasonal product names or a changed form layout will lower the measured rate. re-run the labelled test set whenever the input format changes, or at a fixed interval agreed in the contract.
How much does it cost to run an AI feature in production?
The running cost of an AI feature is the sum of model calls, the human review time it fails to remove, and the engineering time to keep it working. The model call is usually the smallest and most visible line.
AI API cost control starts with the unit of work. If each invoice triggers one call with a fixed prompt and a bounded output, you can calculate cost per invoice. If the feature chains several calls or retries on failure, cost per invoice varies with input quality.
- Set a maximum token budget per request and reject inputs that exceed it rather than sending them.
- Cache results for identical inputs where the business allows it.
- Route routine inputs to a smaller model and escalate only ambiguous cases.
- Log cost per transaction so a price change from the provider is visible immediately.
Human review is the line that surprises buyers. If the feature reaches 90% field accuracy and every record still needs a check, you have added a step rather than removed one. the feature only saves time when the review step is skipped for a defined subset of inputs.
Ask for a cost model expressed per transaction, with the assumptions written down. A provider price change or a rise in input length should show up as a recalculated number, not a surprise invoice.
Custom AI solutions for business are usually priced as a build plus a running estimate. The build is bounded by the test set and the integration points. The running estimate depends on volume, which the client controls.
What should the fallback do when the model is wrong or unavailable?
The fallback should return the user to the last known working step with the input preserved. A graceful failure is one where the person does not have to retype anything and the customer does not see an error.
Design three failure states explicitly: the model returns a low-confidence answer, the model call times out, and the provider is unavailable. Each needs a different response.
A low-confidence answer should route to human review with the source text visible. A timeout should retry once, then fall back. A provider outage should switch to the manual path automatically.
a timeout longer than the user's patience is a failure even if the answer eventually arrives. Set the timeout from the current step's acceptable delay, not from the model's maximum response time.
For a customer-facing feature in Amsterdam, expectation adds a constraint. Customers here are used to same-day email replies and clear status pages. A silent failure that leaves a message unanswered overnight is worse than a slower manual reply.
Write the fallback into the acceptance test. If the test plan only covers successful model responses, the feature has not been specified. Our software development practice treats the failure path as part of the deliverable, and we describe the same pattern for businesses in Amsterdam.
Frequently asked questions
Where does AI actually help a small business?
AI helps most where a person reads unstructured input and types a structured result into a system. Examples are invoice field extraction, delivery confirmation from email, and drafting a reply from a knowledge base. The value comes from removing the transcription step while a person keeps the judgement.
How do you integrate an LLM into an existing system?
Run the LLM beside the current process in shadow mode until its output matches the human output on a labelled sample. Then promote it to a suggestion a person accepts, and only later to an automatic path for narrow cases. The existing manual step stays reachable as a switch throughout.
How much does it cost to run an AI feature in production?
Running cost is model calls plus human review time plus engineering maintenance. Model calls are usually the smallest line and depend on input length and retry policy. Human review dominates unless the feature skips review for a defined subset of inputs.
How do you know if an AI feature is accurate enough?
Measure field-level accuracy on a labelled set of real historical inputs, then compare the error rate to the cost of a wrong answer in that step. A wrong total needs a higher threshold than a wrong internal note. Re-run the test set when input formats change.
What should an AI feature do when the model is unavailable?
It should return the user to the last working step with the input preserved, so nothing is retyped and the customer sees no error. A timeout should retry once and then fall back. A provider outage should switch to the manual path automatically.
Questions we are asked
Where does AI actually help a small business?
AI helps most where a person reads unstructured input and types a structured result into a system. Examples are invoice field extraction, delivery confirmation from email, and drafting a reply from a knowledge base. The value comes from removing the transcription step while a person keeps the judgement.
How do you integrate an LLM into an existing system?
Run the LLM beside the current process in shadow mode until its output matches the human output on a labelled sample. Then promote it to a suggestion a person accepts, and only later to an automatic path for narrow cases. The existing manual step stays reachable as a switch throughout.
How much does it cost to run an AI feature in production?
Running cost is model calls plus human review time plus engineering maintenance. Model calls are usually the smallest line and depend on input length and retry policy. Human review dominates unless the feature skips review for a defined subset of inputs.
How do you know if an AI feature is accurate enough?
Measure field-level accuracy on a labelled set of real historical inputs, then compare the error rate to the cost of a wrong answer in that step. A wrong total needs a higher threshold than a wrong internal note. Re-run the test set when input formats change.
What should an AI feature do when the model is unavailable?
It should return the user to the last working step with the input preserved, so nothing is retyped and the customer sees no error. A timeout should retry once and then fall back. A provider outage should switch to the manual path automatically.
