Once the first agents are up and running, one question always comes up sooner or later: is it really worth it? The general impression is often positive, but an impression is not enough to decide whether to keep an agent, add others or move to a higher plan.
The good news is that you do not need to build a complex dashboard. For a small business, five families of indicators are enough, provided you follow one rule: measure before launching the agent, so you have a point of comparison.
The golden rule: a starting point#
Without an initial measurement, any assessment relies on memory, which flatters or darkens depending on your mood. Before launching an agent, note for one or two weeks:
- the time you spend on the task;
- how many times it is done late or forgotten;
- the average delay between the event and the action (for example between an invoice's due date and the reminder);
- the errors you notice.
An honest estimate is better than nothing. A simple table is enough:
| Task | Time per week | Oversights or delays per month | Usual lead time | Errors noticed |
|---|---|---|---|---|
| Quote follow-ups | ||||
| Meeting minutes | ||||
| Filing invoices |
Family 1: time saved#
What you measure#
The time you used to spend on the task, minus the time you now spend checking and correcting the agent's work.
How to measure it#
For one week, note the time spent reviewing, correcting or completing the agent's results. Compare with your starting point.
The trap#
Forgetting supervision time. An agent that saves you two hours but makes you spend one correcting only saves you one hour. That is still good, but it is the real figure. Supervision time usually decreases after the first few weeks, as you add rules.
Family 2: reliability#
What you measure#
The share of results usable without edits. A simple rate: of the last twenty results, how many did you use as they were, how many did you correct, how many did you throw away?
Indicative thresholds#
There is no universal threshold, but a reading grid can help:
| Results usable as they are | What it means |
|---|---|
| Almost all | The agent is reliable; you can check it less often |
| A large majority | Good level; look at the corrected cases to add rules |
| About half | The brief lacks precision or the task is too ambiguous |
| A minority | Rethink the task: break it down or choose another |
How to improve it#
For each corrected result, note the reason: missing information, wrong tone, unforeseen exception, wrong input data. The recurring reasons point to the rule to add. The article on mistakes to avoid when automating details the most common fixes.
Family 3: lead times#
What you measure#
The time between an event and the corresponding action. Examples:
- time between an invoice's due date and the first reminder;
- time between the end of a meeting and sending the minutes;
- time between receiving a customer request and the first reply;
- time between a job and issuing the invoice.
Why it is often the most telling indicator#
Agents do not necessarily do things "better" than an attentive person, but they do them on time, every time. For tasks such as follow-ups or invoicing, shorter lead times have a direct effect on cash flow and customer satisfaction. A template like the invoice after the job acts precisely on that lead time.
Family 4: oversights avoided#
What you measure#
The number of tasks that would not have been done without the agent: forgotten follow-ups, missed deadlines, unanswered requests, ignored customer reviews.
How to measure it#
It is the hardest indicator to measure, because an avoided oversight leaves no trace. Two approaches:
- compare with the starting point: how many oversights per month before, how many now;
- read the run log: each of the agent's actions is one that used to depend on your memory.
For some agents, such as unanswered emails or contract renewal dates, this is even their main reason for existing.
Family 5: cost relative to usefulness#
What you measure#
The agent's monthly price, compared with the value of the time saved and the effects achieved.
How to calculate it simply#
- Take the time saved per month (family 1).
- Give it a reasonable hourly value: the fully loaded hourly cost of the person who did the task or, for an owner, the value of an hour spent selling or producing.
- Compare with the agent's monthly price.
A fictitious example, to illustrate the calculation: an agent at €19/month that saves two hours a month for someone whose time is valued at €30 an hour produces about €60 of value. It pays for itself on time alone, without counting shorter lead times and avoided oversights.
On Future Agent, each agent's price is set at launch and shown in your dashboard, along with the number of runs and the duration of each. The Pro and Studio plans include several agents and lower the price of additional ones, which changes the calculation once you have several.
Operational signals to watch#
In addition to these five families, a few technical signals are worth a regular look:
- failed runs: an agent that fails often points to a connection problem (a disconnected tool, a changed password) or unexpected input data;
- runs with no result: a watch agent that never finds anything may be poorly targeted;
- run duration: a duration that keeps growing can indicate an increasing volume of data to process;
- storage space: files, tables and sites count towards your plan's storage (1, 25 or 200 GB depending on the plan).
The Future Agent dashboard shows this information for each agent, along with the detailed log of each run.
A fifteen-minute monthly review#
Once a month, go through your agents with this grid:
| Agent | Time saved | Reliability | Lead times | Oversights avoided | Decision |
|---|---|---|---|---|---|
| Keep / Adjust / Stop |
Three possible decisions:
- keep the agent as it is if it does its job;
- adjust its description if reliability or scope can be improved;
- stop the agent if it is no longer useful or the task has disappeared. A stopped agent costs nothing.
You can even automate part of this review: an agent such as the weekly management report or the monthly dashboard can gather your activity indicators in a single document.
An example of a full review#
Here is what the monthly review might look like for a five-person consulting firm using four agents. The observations are illustrative:
- Assignment reports: almost all go out after a quick review; consultants no longer write reports from scratch. Decision: keep.
- Follow-up on sent proposals: follow-ups all go out on time, but two clients found the second message too pushy. Decision: adjust the tone of the second stage.
- Industry watch: the notes are read, but one source keeps coming up and adds little. Decision: adjust the list of sources.
- Meeting preparation: little used, because the consultants prepare their meetings themselves. Decision: stop it and try again in a few months with a different brief.
Fifteen minutes, four decisions, and a setup that improves every month instead of piling up.
What not to measure#
Some indicators look appealing but say very little:
- the number of active agents: ten useless agents are worth less than one agent that changes your week;
- the number of runs: an agent that runs every hour without producing anything useful is not performing;
- the length of deliverables: a two-page report that gets read beats a ten-page report that gets ignored.
Measuring indirect effects#
Some benefits escape the numbers but are worth noting during the monthly review:
- reduced mental load ("I don't think about it any more");
- a better image with customers (quick replies, regular follow-ups);
- continuity during holidays or absences;
- documented processes, which make it easier to onboard a new colleague.
Simply ask each person involved: "If we stopped this agent tomorrow, what would you miss?" The answer is often more revealing than a table.
In short#
- Measure the task before launching the agent.
- Track five families: time saved, reliability, lead times, oversights avoided, cost relative to usefulness.
- Watch for failures and runs with no result.
- Hold a fifteen-minute monthly review with one decision per agent.
- Note the indirect effects by asking the team one simple question.
If you are just starting, the 30-day plan to get started with AI builds this measurement in from the very first week.