The build usually works on demo day. AI projects fail in month four, five, six, when the system meets reality and nobody in the building can fix it. The vendor has moved on, a document format changed, and the tool that impressed the partners in January is quietly abandoned by June.
of enterprise generative AI pilots showed no measurable P&L impact
MIT NANDA, The GenAI Divide: State of AI in Business 2025
If you run a Singapore SME and you are pricing an AI project, this failure pattern is the most important thing to understand before you sign anything. It determines what you should buy, from whom, and what the real cost is.
Where AI projects actually die
Three causes account for most failures.
Hype-led scoping
The project chases what is impressive rather than what pays: a chatbot nobody asked for while four staff spend Thursdays re-keying invoices. This one is fixable with a proper assessment. What that assessment should contain, and cost, is covered in what AI consulting costs in Singapore.
Edge-case builds
The system handles the happy path shown in the demo. The 15% of cases that are messy, the client who sends photos of receipts, the confirmation letter in Mandarin, were never in scope. Staff hit the edge cases, lose trust, and quietly go back to the old way.
No maintenance
This is the big one. AI systems are not spreadsheets. Models get updated by their vendors. APIs change. Your firm changes: new client types, a new document format, a staff reshuffle. A system nobody maintains degrades in weeks and dies in months. The build was fine. The ownership was missing.
Vendors know this, which is why the standard agency answer is a mandatory maintenance contract attached to every implementation.
The retainer math in the Singapore market
The maintenance retainer is not a scam. The maintenance need is real. The question is who does the maintaining.
Run the numbers on a typical done-for-you deal in the Singapore market, at the S$1,500 to S$4,500 monthly retainers commonly quoted:
- System build
- S$18,000
- Retainer, S$3,000 a month for 36 months
- S$108,000
- Three-year total
- S$126,000
A S$18,000 system costs S$126,000 over three years, and at the end of it you own nothing you can operate. Cancel the retainer and the six-month problem starts the day the vendor’s access ends. That is not a maintenance plan. That is a dependency with an invoice schedule.
There is a second, quieter dependency: where the system lives. Some agencies build in their own accounts and environments. If the relationship ends, the system ends. Before any engagement, ask whose workspace, whose API keys, whose logins. If the answer is not “yours,” you are renting.
The alternative: capability transfer
There is a different way to solve the maintenance problem. Make your own staff the maintainers.
Under this model the engagement has three parts, and the third is not optional.
- Diagnose: find the one workflow where the hours and errors concentrate.
- Build: construct the system in your environment, on accounts you own.
- Teach: train the two or three staff who run the workflow to operate, adjust, and repair the system. Written runbooks. Real handover. The capability stays after the consultant leaves.
Compare the three-year cost. A S$35,000 assess-build-train engagement, roughly S$17,500 out of pocket after EDG support at up to 50% for qualifying SMEs, with maintenance done in-house at perhaps 2 to 3 staff-hours a month:
| Done-for-you plus retainer | Build plus train | |
|---|---|---|
| Year 1 | S$54,000 | S$35,000 (S$17,500 post-EDG) |
| Years 2–3 | S$72,000 | ~S$3,000 in staff time |
| 3-year total | S$126,000 | ~S$38,000 gross, ~S$20,500 post-grant |
| End state | Dependency | Owned capability |
Representative arithmetic, conservative on the staff-time side. The gap widens every year the system runs. And the second project is cheaper than the first, because your staff now understand how these systems work. That is the opposite of the retainer curve, where every new system adds another monthly line.
”But my staff aren’t technical”
The most common objection, and the most overweighted.
Modern AI systems for SME workflows are not codebases. They are configurations: prompts, routing rules, document templates, connections between the tools your firm already uses. The person best placed to maintain the invoice-processing system is the person who understands invoices, taught the system. Not the person who understands servers, taught your business.
A well-built system is written the way you would brief a capable new hire: plain language, explicit steps, examples of good and bad output. If the system was built that way, it can be maintained that way. By your people.
What the training must actually cover, concretely: how to read the system’s logs when something looks wrong, how to adjust the instructions when a new document type appears, how to test a change before it goes live, and when to call for help. That is a two-to-three week handover discipline, not a computer science degree.
One honest caveat. Capability transfer requires staff who will engage, typically the workflow’s current owner plus one backup. A firm unwilling to commit those two people to training should buy the retainer model with open eyes. It is the right product for that situation. It is simply not ownership.
What handover should look like
This is the standard to hold any vendor to, including us.
- Built in your workspace from day one. No migration at the end, no hostage infrastructure.
- A walkthrough recording and written documentation, so the system can be understood by a staff member who joins next year.
- An acceptance window where your team runs the system live and every defect found is fixed at no cost. A 30-day warranty is a reasonable market norm.
- Named internal owners, trained hands-on, who have each broken and repaired the system at least once in a controlled setting before sign-off.
- A support option, not a support requirement. Access to help on your terms is healthy. A contract you cannot leave without the system dying is the six-month problem, deferred and compounding.
A vendor who resists this list is telling you their revenue model. Listen.
How Kept structures this
Kept’s engagement is the capability-transfer model priced plainly. The diagnostic is 2 to 3 weeks at S$5,000 to S$15,000 by firm size. Build & Train is 6 to 12 weeks at S$25,000 to S$45,000, fixed scope, built on your tenant, with staff training and a 30-day warranty inside the scope. A care plan at S$1,500 to S$3,000 a month exists as an option, never as a condition of the system surviving. Custom projects of this shape can qualify for EDG support of up to 50%, subject to EnterpriseSG approval; mechanics are on enterprisesg.gov.sg and applications run through the Business Grants Portal.
The vertical version of this argument, with audit-file and month-end arithmetic, is in AI for accounting firms in Singapore.
Common questions
Why do most AI projects fail?
Three dominant causes: hype-led scoping that automates the wrong thing, builds that ignore messy edge cases so staff abandon the tool, and, most often, no maintenance after handover. AI systems degrade as models, APIs, and the business change. MIT's NANDA research group reported in 2025 that 95% of enterprise generative AI pilots showed no measurable P&L impact, overwhelmingly for post-launch reasons.
How do I make a long-term AI project succeed?
Someone has to maintain it. Either pay a vendor a monthly retainer indefinitely, or have the engagement train your own staff to operate and repair the system. Insist the build lives in your accounts, comes with documentation and a defect-warranty window, and names internal owners before sign-off.
Are AI maintenance retainers worth it?
The maintenance need is real; the retainer is one way to meet it. At the S$1,500 to S$4,500 a month typical in Singapore, a retainer adds S$54,000 to S$162,000 over three years to a project's cost. If your staff can be trained to run the system, usually 2 to 3 hours a month of internal effort, the retainer solves a problem you no longer have.
What is capability transfer in AI consulting?
An engagement model where the consultant builds the system and then trains the client's staff to run it, so the capability stays in the firm after the engagement ends. It replaces the ongoing retainer with a one-time training investment and written runbooks: ownership instead of dependency.
