The Last Mile of Coaching: Why Good Data Fails at the Final Step

A parcel can cross an ocean, clear customs, pass through three sorting hubs and arrive at the right street on the right day, and still be left in the rain. Logistics has a name for the leg where that happens. The last mile is the shortest part of the journey and the most expensive: Capgemini's research puts it at 41 per cent of total supply chain cost, more than double any other stage, and the stage where most of the complaints are generated (Capgemini Research Institute, 2019). Everything upstream can be run beautifully and the customer only remembers the doorstep.

Coaching in a contact centre has a last mile too, and it is where the money goes missing.

What is the last mile of coaching?

The last mile of coaching is the final step between a performance signal and a change in behaviour: the moment a team leader or QA coach turns what the data says into words a particular person hears. Call recording, speech analytics and QA scoring read the call; the last mile is how the finding is delivered to the agent who was on it. The largest review of feedback ever run found more than a third of feedback interventions made performance worse, so this step decides whether the data changes anything.

The engine room has been solved

For most of the sector's history, the problem with coaching was that nobody had seen the call. Manual quality assurance sampled between 1 and 3 per cent of interactions, scored them against a rubric, and met the agent about it the following week (Verint, 2026). Ninety-seven calls in a hundred went unheard. A team leader coached from memory, from a handful of calls, and from whatever the agent chose to mention.

That problem is now, more or less, solved. Speech analytics and conversation-intelligence tools, RingSense and Gong among the better known, along with the automated QA that Verint, NICE and others sell into contact centres, can score every interaction against the campaign standard, flag a missed disclosure or a competitor mention left out of a pitch, and do it the same way every time. The signal is complete and the signal is consistent. If the engine room was the problem, the problem has gone.

And yet the sector's engagement and attrition numbers have not moved in step with its analytics budgets. Gallup records 20 per cent of employees engaged worldwide and managers as the least engaged group it measures, at 22 per cent (Gallup, 2026). COPC still puts around 70 per cent of a centre's first-year attrition inside the first ninety days (COPC, 2024). Better data about the call has not, on its own, produced better conversations about the call. Something between the dashboard and the person is not working, and the research on feedback has known what it is for thirty years.

The evidence that delivery is where it fails

Does more feedback improve performance?

Not reliably. Kluger and DeNisi's meta-analysis of 607 effect sizes found feedback improved performance on average but made it worse in over 38 per cent of cases, and the harm was not explained by whether the feedback was positive or negative. Gartner found no significant relationship between the time a manager spends coaching and employee performance; managers who personalise coaching to the person lifted performance by up to 26 per cent while always-on managers lowered it by up to 8 per cent.

In 1996 Avraham Kluger and Angelo DeNisi pulled together every usable study of feedback and performance they could find, 131 of them, 607 effect sizes, 23,663 observations. The average effect was positive, a respectable d of 0.41. The finding that made the paper a classic sat underneath the average: in over 38 per cent of cases, feedback made performance worse (Kluger and DeNisi, 1996). Not no effect. Worse. And the sign of the feedback did not explain it; praise could harm and criticism could help. What explained it, as far as their moderator analysis could tell, was where the feedback pointed the recipient's attention. Feedback that kept attention on the task tended to work. Feedback that moved attention up the hierarchy to the self, to how the person was being judged, tended not to. I wrote about this study in the review of the feedback sandwich; the point here is simpler. The same finding, delivered two ways, can produce opposite results, and the difference is entirely in the delivery.

Gartner's manager research reached the same conclusion from the other side. In a multi-year study of thousands of employees and managers, the firm sorted managers into four approaches and measured what happened to their people. The "always-on" manager, the one who gives frequent feedback across every skill and is the approach most organisations say they want, lowered employee performance by up to 8 per cent. The "connector" manager, who gets to know the person, coaches where they have the expertise and connects the employee to someone better placed where they do not, raised it by up to 26 per cent (CIO Dive, 2019, reporting Gartner). Employees who report to effective coaches are 40 per cent more engaged, show 38 per cent more discretionary effort and are 20 per cent more likely to stay (Gartner). The first thing Gartner lists under what connectors do is "personalise for employee resonance": tailor the development to the person.

Gallup has measured the last mile from the employee's side. Only 16 per cent of employees say their last conversation with their manager was meaningful, while 80 per cent of those who received meaningful feedback in the past week are fully engaged, and what makes a conversation meaningful, in the order people rank it, is appreciation for recent work, the relationship, current goals and the person's strengths, with weaknesses rated least (Gallup, 2026). Among employees who strongly agree their manager focuses on their strengths, 67 per cent are engaged; among those who strongly disagree, 2 per cent are (Gallup). The dashboard, by design, is a list of what went wrong on the call. Handed over as it stands, it is the conversation people rate least.

The same finding, delivered two ways, can produce opposite results, and the difference is entirely in the delivery.

Why the same note lands differently

Hold Kluger and DeNisi's mechanism against a real coaching floor. A team leader has a roster of twelve or fifteen agents, a few minutes for each, and a scorecard that says what the call needed. Out of that comes a note, and the note is accurate and uniform. The same note lands on the agent who wants the exact timestamp and the evidence first, and on the agent who needs to understand the point of the standard before any detail will stick. It lands on the agent who will work the problem through alone if given a day, and on the agent who wants to talk it out now. For one of them it stays on the task. For another it goes straight to the self: I am being watched, I am being marked, this is about me. The note was the same. The last mile was not.

These are not personality types and they are not skills. They are work preferences, the ways a person is naturally inclined to connect, think, decide and deliver, and I have set out what they are and are not in a separate piece. What matters here is that a preference is exactly the kind of information the dashboard does not hold and the team leader rarely has time to discover for fifteen people in the first ninety days. The analytics know the call. Nobody has told the team leader about the person.

This is the gap the logistics industry would recognise at once. The sorting hub is not the problem. The parcel is addressed correctly. The question is what happens at the door, and the industry's answer, after a decade of trying to automate the doorstep, has been to get better at knowing the recipient: when they are in, where they want it left, who to call.

What is a preference layer in coaching?

A preference layer sits between the data about a call and the conversation about it. It holds how each agent prefers to work and to receive feedback, read from a short survey, and uses it to shape a coaching note: whether to lead with the evidence or the point, whether to ask an open question or walk through the fix, how to open with this person. The recording, analytics and QA scorecard stay as they are; the preference layer changes only how the finding is delivered.

One compliance flag, two routes to the agent

Here is how it plays out, with Alex, the agent who runs through the book behind Sariio. The call is a real practice call, read against the campaign's own 30-point QA sheet; Alex's work preferences are real and her name is not. At 1:50 Alex told the customer she was adding a refund to their account. At 2:15 her closing summary described the resolution as a discount on their next trip, and the customer confirmed the inaccurate version. Under the campaign's scorecard that is a fail on the closing summary, regardless of how good the rest of the call was, and the rest of the call was good: the structure held from introduction to close, the customer was named at the end, and the sheet came back at 76 per cent. The automated QA flags the summary. From this point the data is identical on both routes. Only the last mile differs.

The scorecard figures and the one2one are those of Alex's one2one; the standard route is drawn from the generic coaching templates that ship with QA and conversation-intelligence tools.

Nothing on the second route required a new data source, a new scorecard or a longer conversation. It required knowing one thing about the person that the call could not tell anyone, and letting that one thing shape five minutes. That is the whole of the last mile, and it is what turns a team leader from a data-checker into a coach.

What happens when two agents take the same call?

When two agents are coached on the same call, the data does not move and the coaching should. In a live test, one practice call was run through Sariio twice against two agents' myMAP profiles. All 30 QA verdicts were identical, 76 per cent both times, and both notes chose the same moment, a refund offered at 1:50 described as a discount at 2:15. The one2ones differed in what they opened with, how the miss was framed, what to avoid and the habit each agent was asked to try. Same call, same score, different last mile.

One real practice call, run through Sariio twice against two agents' myMAP profiles. Names are changed; the verdicts, timestamps and coaching text are as produced.

The virtuous circle, when coaching lands

How does coaching affect retention and performance?

When coaching lands the effects compound. A field study of 5,172 support agents found AI-prepared assistance lifted issues resolved per hour by 13.8 per cent and cut turnover by 8.6 points. Gallup finds 52 per cent of voluntary leavers say their exit was preventable. Peer-effects research shows a 10 per cent rise in co-worker productivity lifts a colleague's own by around 1.7 per cent, and Gallup's meta-analysis puts top-quartile engaged units 23 per cent ahead on profitability with 21 to 51 per cent lower turnover.

Coaching that lands improves the next call; people whose performance improves are more likely to stay; people who stay and succeed influence the colleagues who can see them; and the sum of that is retention, culture and commercial performance, which in turn buys the team leader the trust and the time to coach again. Each link in that loop has evidence behind it.

Coaching that lands improves performance. The cleanest field evidence comes from Brynjolfsson, Li and Raymond, who tracked 5,172 support agents as a generative AI assistant was introduced behind them. Issues resolved per hour rose 13.8 per cent on average and around 35 per cent for the newest agents; the tool was, in effect, coaching delivered in the moment, in a form the agent could use (Quarterly Journal of Economics, 2025). Gartner's connector finding sits alongside it: coaching tailored to the person lifted performance by up to 26 per cent while constant generic feedback lowered it.

People whose performance improves are more likely to stay. In the same study agent turnover fell by 8.6 points, and the fall was concentrated among the newer agents who had improved most. Gallup's exit research says why that matters: 52 per cent of people who left voluntarily say their manager or organisation could have done something to keep them, and 51 per cent say that in their last three months nobody spoke to them about their future (Gallup, 2019). For an agent who leaves at week ten, those three months were the whole employment, and the conversations that did or did not land were the whole relationship.

People who stay and succeed lift the people around them. This is the link most coaching business cases leave out, and it has some of the best evidence. Mas and Moretti measured the output of supermarket cashiers minute by minute and found that when a highly productive colleague joined a shift, the others sped up: a 10 per cent rise in average co-worker productivity lifted a worker's own by about 1.7 per cent. The effect ran only through colleagues who could see each other, and it was strongest among the workers who had been below average (Mas and Moretti, 2009). A meta-analysis of 34 such studies puts the average spillover at 1.2 per cent for every 10 per cent (Herbst and Mas, 2015, via Economic Inquiry). A contact-centre floor, where everyone can see and hear everyone, is close to the ideal conditions for it. The mood travels as well as the pace: Sigal Barsade's experiments showed that a single group member's positive affect spread to the others and improved cooperation, reduced conflict and raised the group's sense of its own performance; people, in her phrase, are "walking mood inductors" (Barsade, 2002). A can-do agent is contagious, and so is the opposite.

The sum is retention, culture and commercial performance. Gallup's eleventh Q12 meta-analysis, across 183,806 business units and 3.3 million employees, puts the median gap between top- and bottom-quartile engaged units at 23 per cent in profitability, 18 per cent in sales productivity, 10 per cent in customer loyalty, 78 per cent in absenteeism and between 21 and 51 per cent in turnover, depending on how much turnover the organisation already carries (Gallup, 2024). Culture, as I have argued in another piece, is the collective language and behaviour of an organisation's people; a floor where the coaching lands and the agents stay long enough to shape one another is a culture being built one conversation at a time.

And the circle closes. A team leader whose coaching has landed a few times has a floor that trusts the conversation, agents who arrive with their own fixes, and the minutes that used to go on preparation and on backfilling leavers. That is the time and the standing to do the next last mile well.

Nothing on the second route required a new data source, a new scorecard or a longer conversation. It required knowing one thing about the person that the call could not tell anyone.

Where Sariio sits

Sariio is built for the last mile and for nothing upstream of it. The recording, the analytics and the QA scoresheet stay exactly as they are; they are good at what they do and we have no wish to replace them. What Sariio adds is the preference layer: a map of how each agent prefers to work and to receive feedback, read from a twelve-minute survey the agent sees before anyone else does, sitting beside whatever the operation already produces about the call.

From the flag, the campaign standard and the agent's myMAP, Sariio prepares the one2one, a short note for the team leader that quotes what went well, names the moment worth revisiting alongside the standard, proposes one action to adopt, and suggests how to open with this particular person. The team leader reads it, changes what needs changing and endorses it; only then does the agent see anything. The judgement and the conversation remain the team leader's. What changes is that the team leader walks in knowing the person, which is where, by Gallup's account, meaningful conversations begin. Between one2ones the agent carries on with Sariio Coach about the action they chose, and the team leader can ask how best to open the next conversation. The book behind Sariio follows Alex through exactly this; the first spreads are online, and her sample one2one is the one used above.

Across a department the same preference data reads as a cultureMAP, which is the subject of a separate piece. For this one, the point is narrower. The sector has spent a decade and a great deal of money on the sorting hubs. The parcel now arrives at the right door every time. The last mile is where it is decided whether anyone opens it.

---

The companion pieces: the feedback evidence in full is in the feedback sandwich, reviewed; the case for AI that prepares rather than replaces is in the AI paradox; the first ninety days are in why new call centre agents leave so soon; and what a work preference is, and is not, is in what are work preferences.

---

Sources

• Capgemini Research Institute (2019), 'The last-mile delivery challenge' - last-mile services at 41% of overall supply chain cost, more than double any other category; 2,874 consumers and 500 executives surveyed
• Verint (2026), 'Call center quality monitoring: best practices for QA' - manual QA sampling 1 to 3 per cent of interactions; automated scoring of up to 100 per cent
• Kluger, A. N. and DeNisi, A. (1996), 'The effects of feedback interventions on performance: a historical review, a meta-analysis, and a preliminary feedback intervention theory', Psychological Bulletin 119(2) - 607 effect sizes, 23,663 observations, d = 0.41, over one third of interventions negative; attention moving towards the self reduces effectiveness
• CIO Dive (2019), 'Gartner: of the 4 manager types, only 1 boosts employee performance 26%' - reporting Gartner's connector manager research; always-on managers degrade performance, connectors lift it by up to 26%
• Gartner, 'Connector Manager' executive guidance - employees of effective manager coaches 40% more engaged, 38% more discretionary effort, 20% more likely to stay
• HBR IdeaCast (2023), 'The best managers are connectors' - Sari Wilde of Gartner on finding no significant relationship between time spent coaching and employee performance
• Gallup (2026), 'The great manager's most important habit' - 80% and 16%; the ranking of what makes feedback meaningful
• Gallup, 'Global study: ROI for strengths-based development' - 67% engaged when the manager focuses on strengths, 2% when they strongly disagree
• Gallup (2026), 'State of the Global Workplace 2026' - 20% of employees and 22% of managers engaged
• Gallup (2019), 'This fixable problem costs US businesses $1 trillion' - 52% of voluntary leavers say it was preventable; 51% had no conversation about their future in their last three months
• Gallup (2024), 'Q12 Meta-Analysis, 11th edition' - 183,806 business units, 3.3 million employees; 23% profitability, 18% sales productivity, 10% customer loyalty, 78% absenteeism, 21% and 51% turnover
• COPC (2024), 'Onboarding best practices: reducing attrition in the first 90 days' - around 70% of first-year attrition inside 90 days
• Brynjolfsson, E., Li, D. and Raymond, L. (2025), 'Generative AI at Work', Quarterly Journal of Economics 140(2) - 5,172 agents; 13.8% more issues resolved per hour, around 35% for novices, turnover down 8.6 points
• Mas, A. and Moretti, E. (2009), 'Peers at Work', American Economic Review 99(1), NBER Working Paper 12508 - a 10% rise in average co-worker productivity associated with a 1.7% rise in a worker's own effort; effect runs through co-workers who can see the worker
• Herbst, D. and Mas, A. (2015), via Economic Inquiry (2024) - meta-analysis of 34 studies; a 10% rise in co-worker productivity lifts own productivity by 1.2% on average
• Barsade, S. G. (2002), 'The ripple effect: emotional contagion and its influence on group behavior', Administrative Science Quarterly 47(4) - positive contagion improved cooperation, reduced conflict and raised perceived task performance