Most AI robot projects fail not because of the robot but because of the job chosen for it. That choice is made before any product is evaluated, and it matters more than every specification comparison that follows.
Five criteria for a first job
It repeats many times a day. A task occurring a few times a week leaves the robot idle, and idle is the fastest route to forgotten.
The content is stable. Answers do not change daily. If information shifts constantly, maintenance cost exceeds the benefit.
Errors are harmless. A robot giving wrong directions mildly inconveniences a visitor; a robot quoting a wrong price costs money. A first project belongs in the first category.
It is measurable. Visitors served, questions handled, deliveries completed. Without a number, nobody knows after three months whether it was worth continuing.
There is a manual fallback. The robot will charge, be serviced, and one day fail. A job that stops entirely without it is not a first job.
The tasks meeting all five in most businesses sit in the customer-facing area rather than in production.
Three task types that fit best
Answering repeated questions. At any reception desk a large share of questions are the same ones. The robot absorbs those; staff handle the rest. Value is immediately measurable in questions never reaching a person.
Guiding and presenting along a route. Showrooms, facilities, buildings, property developments. Fixed content, fixed route, and every visitor receives the full presentation — something a human guide struggles to sustain after the twentieth repetition.
Covering periods with no staff. After hours, peak times when everyone is occupied, small branches without a permanent presence. Here the comparison baseline is nothing, so any service is an improvement and no one feels threatened.
These three share a property: the core work is talking and moving, which is exactly the strong half of current AI robot capability, and none of them touches hand manipulation, which is the weak half.
What to avoid on a first project
Tasks needing hand manipulation of varied objects. Clearing tables, stocking shelves, preparing drinks. Appealing, and not reliable outside demonstrations.
Tasks in dense crowds. Markets, packed events, food courts at peak. Moving well in a corridor does not mean coping in a crowd.
Tasks where one wrong sentence has consequences. Legal advice, medical guidance, quoting, commitments to customers. If you proceed anyway, a person must approve before information reaches the customer.
Replacing an entire role. Not technically feasible and it generates internal resistance that damages the project itself.
Tasks where content changes daily. Update cost will exceed the value delivered.
The general principle for a first project: choose a job where a poor robot performance costs nobody anything significant. That is what makes learning affordable.
Finding the job inside your own company
A short process that fits in two weeks.
One: log every question asked. For one week, reception staff write down every question visitors ask. Count the repeats at the end. This data is valuable and costs almost nothing to collect.
Two: measure walking time. How many minutes a day staff spend guiding visitors, carrying items, fetching documents.
Three: identify the gaps. Which hours and which locations currently have nobody serving visitors.
Four: ask the staff. Which tasks feel like wasted time. The answers are usually blunt and useful.
Five: pick one job. Just one. The highest-frequency, lowest-risk item on the list.
This produces a decision based on your own data rather than on a vendor's suggestion, and that is the largest single difference between projects that produce results and projects bought on enthusiasm.
Writing success criteria before starting
Half a day of work that shapes the whole project.
Write the numbers. After three months, how many interactions per day, what proportion answered correctly, how many human interventions per day.
Include interim milestones. Month one expectations should be far below month three, because month one is for learning and adjusting.
Name who measures and how. Without a person and a method, the numbers are never collected.
Define a stopping condition. If it does not reach a given level after three months, pause and review. Writing this beforehand is far easier than deciding under pressure to justify money already spent.
Note the unquantified benefits separately. Visitor impression, brand image, organisational experience. Record them without assigning numbers, so the quantified part keeps its credibility.
Frequently asked questions
What should the first job look like?
Repeating many times a day, with stable content, where errors are harmless, results are measurable, and there is a manual fallback when the robot is unavailable. Tasks meeting all five are usually customer-facing.
Why are the best applications all talking and moving?
Because that is the strong half of current AI robot capability. Hand manipulation of varied objects remains the weak half, so applications touching it fail at high rates despite sounding more impressive.
How do you find the right job internally?
Log every visitor question for a week and count the repeats, measure staff walking time, identify hours with nobody serving visitors, and ask staff directly which tasks feel like wasted time.
Why write success criteria before starting?
Because objectivity becomes much harder once money is spent. Write the target numbers for month three, who measures them, how, and the condition under which the project pauses for review.
More in Deployment and operations and Robots for real estate and showrooms.
