Silver Triangle in the AI Era: Employee, Company, Customer

Chapter 1. Where the triangle comes from

• Company → customer (external marketing): making the promise. Everything the organization communicates, offers, prices and promises, through any channel and any intermediary.
• Company → employee (internal marketing): enabling the promise. Selection, training, tools, information, autonomy and incentives. Without these the promise still gets made, but nobody is in a position to hold it up.
• Employee → customer (interactive marketing): keeping the promise. The actual encounter, which is where the promise is either validated or falls apart.

Chapter 2. Why this is central for a startup today

Chapter 3. The employee – company relationship: enabling the promise

• High-volume recruiting and screening. This is the most mature use of all, and Gartner notes that in 2026 high-volume recruiting (operational roles, retail, customer service) goes outright AI-first, because that is where the savings are greatest and where each individual decision carries least complexity. AI ranks applications, runs a first structured interview, verifies hard requirements and coordinates calendars. The risk is well known and worth anticipating: a model trained on résumés learns to reward the shape of a résumé, which does not always correlate with ability, so the cut-off criterion has to remain an explicit, documented human decision.
• Performance reviews. The annual review carries two flaws we all recognize: it rests on the manager's recent memory, which rarely reaches beyond the last quarter, and it gets written in a rush the week before the meeting. With AI it becomes possible to move from the event to the continuous record, gathering across the year the scattered evidence (objectives, deliveries, feedback received, project participation) and producing a draft the reviewer corrects and signs. The interesting effect is not the writing time saved, but that the annual conversation stops being consumed by arguing about whether things happened, and can be spent on what it should be about, which is where that person is heading.
• Talent discovery and skill inference. Organizations know precisely what position each person holds, but very few know what each person can actually do. Talent intelligence systems infer skills from real activity (projects worked on, tools used, training completed, previous moves) and build an inventory that updates itself. When that inventory is crossed with open needs, people turn up who are capable of things their job description never mentioned — which is more or less the operational definition of hidden talent.
• Internal talent marketplace and mobility. On that same inventory you build the matching between people and internal opportunities, whether short projects, temporary cover, formal vacancies or mentoring. We consider it the best effort-to-return ratio in the whole chapter, because it attacks two problems usually treated separately: the cost of recruiting outside what you already have inside, and one of the most cited reasons for resignation, which is not seeing a path forward within the organization.
• Mismatch between capabilities and needs. Crossing the skills inventory with the business plan lets you see where capacity will be missing before it actually is. Gartner projects that by 2030 half of enterprises will face irreversible skill shortages in at least two critical job roles, partly because of the erosion of competencies in tasks delegated to AI that nobody afterwards knows how to do again. Against that horizon, spotting the mismatch twelve months ahead allows you to train; spotting it once it has happened forces you to go buy talent at the worst moment and the worst price.
• Detection of anomalous situations. Unusual access to sensitive information, approvals that depart from the usual pattern, spending deviations, controls that get skipped. At volume, these are signals no team reviews by hand, and that a model can flag for someone to look at. It is worth being explicit about the boundary before switching the system on rather than after: this is risk control over processes, not surveillance over people, and the difference between the two depends on what is monitored, who sees it and what is done with the result — all of which is defined in writing.
• Early signals of disengagement and attrition. Attrition prediction models combine tenure, relative compensation, internal mobility and survey results. Used well, they let a manager have in time a conversation they would otherwise have too late. Used badly, they degenerate into a list of people «at risk of leaving» that circulates through the organization and ends up causing what it meant to prevent. The rule we recommend is simple: the output goes only to whoever can do something for that person, and it does not go into the personnel file.
• Onboarding and an internal knowledge assistant. An assistant able to answer questions about policies, procedures, systems and who does what, always showing the source of its answer, resolves most of a new hire's questions without occupying a colleague every time. It also leaves behind a by-product that is often worth more than the saving: the record of what people ask and fail to find, which works as a fairly honest map of where the organization documents badly.
• Personalized learning and a shorter experience curve. The work of Brynjolfsson, Li and Raymond, published in 2025 in the Quarterly Journal of Economics on data from 5,172 support agents, measured an average productivity increase of 15% on introducing a conversational assistant — but what matters for this chapter is how that gain is distributed. Less experienced and lower-performing workers improved in both speed and quality, while the most experienced barely gained speed and lost some quality. The mechanism the authors identify is that the model diffuses the practices of the best agents among the rest, which turns the tool into an instrument of training rather than of productivity, and considerably changes the calculation of who is worth hiring and how long it takes to bring them up to speed.
• Administrative load handed back to the relationship. Meeting minutes, project summaries, committee reports, answers to repetitive administrative queries. It is the least glamorous use in the chapter and probably the one with most impact on climate, because what it gives back is conversation time between managers and teams, which is precisely the raw material of this side of the triangle and the first thing sacrificed when the calendar tightens.

Chapter 4. The company – customer relationship: making the promise

• Intake of requests and orders through any channel. An order can arrive by email, through a form on the site, by WhatsApp, by phone, or inside an attached PDF someone scanned crooked. AI normalizes all of that into a structured request, classifies it, detects what is missing and injects it into the right process. For a mid-sized organization it is usually the first use case with a clear return, less for the data entry it saves than for the order that stops getting lost in an inbox nobody checks on Fridays. It is also the point where this side of the triangle connects with business process management.
• Autonomous resolution of queries. Gartner projects that by 2029 agentic AI will autonomously resolve 80% of common customer service issues, with a reduction of around 30% in the area's operating costs. The report itself adds a nuance that sometimes gets lost in the citation: human roles do not disappear, they shift toward complex cases and toward a new function, which is supervising and managing the automated agents.
• Personalization and recommendation. Ordering the catalogue, the communication and the offer according to each customer's behaviour is the oldest use of analytics applied to marketing, and yet it is among those that gained most from current models, because they now take in free text. What a customer wrote in a complaint or a review weighs as much as what they bought, and that information simply did not enter the model before.
• Demand forecasting and pricing. Models that anticipate volume by product, channel and period and adjust price or availability accordingly. Well calibrated, they avoid both stockouts and clearance sales, which are the two ways of losing margin on the same forecasting error. Badly calibrated, they produce situations where a customer discovers they paid differently from their neighbour for the same thing, which beyond a trust problem is, in several countries, a legal one.
• Churn prediction and retention. Identifying far enough in advance the customers with a high probability of leaving makes sense insofar as something is done with that information. The habitual mistake is to use it only to trigger an automatic discount, which buys time without resolving anything; the more productive use is to understand what pattern of experience systematically precedes departure and correct that pattern, which over the medium term costs less than the discount.
• Voice of the customer without a survey. NPS and its relatives measure a small sample, made up moreover of people willing to answer surveys, which is not a neutral sample. With AI you can instead analyse the totality of real interactions (calls, chats, emails, complaints, public reviews) and extract from them the concrete sources of friction with their frequency and their evolution over time, which is actionable in a way an aggregate index never is.
• Execution and delivery. Dispatch routing, technician assignment, capacity planning, proactive notice when something is going to be late. This is the part where the promise is kept or broken in verifiable terms, with no room for interpretation, and where optimization hits cost and satisfaction at the same time, which is not common.
• Risk, fraud and customer onboarding. Identity verification, credit assessment, detection of anomalous transactions. Here artificial intelligence performs a double function worth keeping in view: it reduces expected loss, of course, but when well calibrated it also shortens onboarding time — and onboarding is the first concrete experience a customer has with the company after having bought the promise.
• Agentic commerce, or the customer who is no longer a person. This is the newest item in the chapter and the least discussed in boardrooms. With protocols such as the Agentic Commerce Protocol, which Stripe developed together with OpenAI, or AP2, which Google announced in September 2025 with more than sixty participating organizations including payment networks, brands and platforms, an AI agent can search, compare, decide and pay on behalf of its user. That means a growing share of the promise is no longer made to a person looking at a page, but to a model that reads structured data and compares. Whoever does not publish catalogue, price, availability and terms in a format an agent can consume risks being left out of the comparison altogether, without ever learning it took place.
• Being found by the models. The previous point has a corollary that applies to content rather than catalogue. As traditional search cedes ground to generated answers, the question a commercial team used to phrase as «how do I rank on Google» is turning into «what does a model answer when asked about my category». The rules are not those of classic SEO, and the discipline is still forming, but it belongs squarely to the external side of the triangle, just as advertising does.

Chapter 5. The employee – customer relationship: keeping the promise

• Drafting communications. Drafts of emails, proposals, replies to complaints, follow-up messages. Speed is the first thing noticed, but the more durable benefit is consistency, because it lets a salesperson six months into the job produce a communication with the tone, structure and level of detail the organization wants, and turns their work into editing rather than inventing from scratch.
• Real-time assistance during the contact. This is exactly what the Quarterly Journal of Economics study cited earlier measured: suggestions that appear while the conversation unfolds. Beyond the 15% average productivity gain, the authors found two effects that almost never make it into the business case used to approve such a tool. The first is that customers become more polite and ask to speak to a supervisor less often. The second is that agent retention increases, driven above all by keeping the newer workers. In other words, this is the point where all three sides of the triangle touch at once.
• Alerts on angry customers. Sentiment and escalation detection while the conversation is happening, with notice to the supervisor before it breaks. The value is in when the alert arrives, because an escalation caught live still admits intervention, whereas the same escalation caught in next week's quality meeting only admits analysis.
• Customer prioritization. Ordering the day's agenda by value, risk, urgency or probability of closing, instead of by order of arrival or — as happens more often than is admitted — by the salesperson's personal affinity with the customer. It looks like a minor change and it is not, because contact time is the scarcest resource on this side of the triangle and tends to be allocated with no explicit criterion at all.
• Preparing for the encounter. A summary that arrives before the meeting or the call, with the account history, the commitments left open, the latest complaints, the customer's actual use of the product and public news about their company. It is the kind of preparation everyone recognizes as necessary and almost nobody does — not out of negligence but because assembling all of it by hand takes forty minutes that never exist.
• Post-contact logging and case closure. Summary of what was discussed, CRM update, committed tasks, follow-up email. Beyond the time it returns, it resolves the chronic problem of commercial systems, which is that information arrives late, arrives incomplete or does not arrive at all — so the system ends up reflecting what salespeople felt like entering rather than what actually happened.
• Language. Real-time translation and adaptation, written and spoken. For a company exporting services from Latin America this widens the addressable market without changing the team, which is why it is worth treating as a commercial decision discussed in committee and not as one more feature of the service software. It is worth remembering, though, that translating the product is not localizing the business.
• Quality and coaching across all interactions. Traditional quality control listened to a handful of calls per agent per month, chosen more or less at random, and built an evaluation on that minimal sample. Today it is possible to evaluate the whole with uniform criteria and return specific feedback on concrete cases. The correct use of this is formative; if instead it is turned into a sanctions dashboard, the team quickly learns to satisfy the automated evaluator rather than the customer, and the interactive side degrades while the indicator improves.
• Next best action. Recommending, in the context of the conversation taking place, what to offer, resolve or escalate. It works well when the recommendation comes with its rationale, because then the person can assess it and decide; it works badly when it arrives as an instruction without explanation, because in that case whoever is handling the contact stops contributing judgement and merely executes, which is exactly the opposite of what this side of the triangle is for.
• Tracking the promises made in the conversation. This is the least implemented use and the one that connects most directly with the model we have been discussing. In every contact concrete commitments are taken on («I'll send you the quote tomorrow», «I'll escalate it to support and get back to you Thursday») that today live in the memory of whoever made them and nowhere else. Extracting them automatically from the conversation, turning them into tasks with due dates and flagging them as they come due amounts, literally, to automating the keeping of the promise, which is the very definition of this side of the triangle.

Chapter 6. Conclusions

Buho Advisors — Technology advisory for boards in Latin America.
Real experience. Strategic vision.