The problem with traditional evaluation

The annual performance evaluation is the most hated HR process in the company. Employees hate it because they feel their year of work is reduced to an arbitrary number. Managers hate it because they have to fill in endless forms with meaningless text. And HR hates it because they know the results are contaminated by bias and rarely translate into concrete actions.

The problems with traditional evaluation are well known:

• Recency effect: the manager remembers what happened in the last 4 weeks, not the last 12 months. A December mistake weighs more than 11 months of good performance.
• Halo effect: a positive (or negative) first impression contaminates the entire evaluation. If the employee is likable, all their competencies are rated higher.
• Central tendency: managers avoid extremes. Everyone is "satisfactory" or "good". Real differentiation disappears.
• Affinity bias: we evaluate people who resemble us better (same university, same hobbies, same communication style).
• Lack of data. The evaluation is based on the manager's memory and impressions, not on objective performance data.

AI does not make evaluation a perfect process. But it can mitigate each of these problems if used correctly.

Automated OKR tracking with AI

OKRs (Objectives and Key Results) are the most widely used goal-setting framework in tech companies and increasingly in other sectors. The concept is simple: you define an Objective (qualitative, inspiring) and 3-5 Key Results (quantitative, measurable) that indicate whether you achieved it.

Where OKRs fail without AI

The problem is not defining OKRs. It is tracking them. In most companies, OKRs are defined in January, superficially reviewed in July, and rediscovered in December when evaluation time comes. By then, half are obsolete because the context has changed.

How AI improves OKR tracking

• Assisted definition. AI helps formulate OKRs that meet SMART criteria. A vague objective like "improve customer service" becomes "Reduce average ticket response time from 24h to 8h in Q2, measured by the ticketing system".
• Automatic updates. If Key Results are linked to internal system data (CRM, project tool, product metrics), AI can update progress automatically. Without anyone having to fill in anything.
• Deviation alerts. If a KR is below 30% at mid-quarter, AI alerts the employee and manager. Not at quarter end when it is too late, but in time to correct.
• Adjustment suggestions. If context has changed (a team member left, a new priority emerged), AI suggests OKR adjustments with justification.

Prompt to define OKRs with AI

"Help me define OKRs for [name/position] for next quarter. Department context: [priorities]. Their role includes: [main responsibilities]. Generate 2-3 Objectives with 3-4 Key Results each. Each Key Result must be: measurable (with metric and numerical target), ambitious but achievable, and linked to a system where it can be verified. Include for each KR: metric, current value (baseline), target, and data source."

360 feedback analysis with AI

360 feedback collects evaluations from multiple sources: direct manager, peers, direct reports, and sometimes internal clients. It is more comprehensive than manager-only evaluation. But it also generates a volume of data that is hard to process.

The 360 problem without AI

A 360 for an employee with 5 evaluators generates 30 to 50 responses (between quantitative and open). For a team of 20 people, that is 600 to 1,000 responses someone has to read, analyze and synthesize. That "someone" is normally the manager, who does not have time and ends up doing a superficial summary.

How AI analyzes 360 feedback

• Synthesis by competency. AI groups responses by competency and generates a summary with key themes, identified strengths and areas for improvement. Includes anonymized text quotes as evidence.
• Discrepancy detection. If the manager rates a competency as excellent but peers rate it as needs improvement, AI flags it. These discrepancies are the most valuable conversations in the process.
• Language analysis. AI detects patterns in open responses: are the same words repeated? Is there a theme all evaluators mention? Is there an evaluator who clearly has a personal bias (everything positive or everything negative)?
• Executive report. For each employee, a 1-2 page document with: strengths (3 bullets), development areas (3 bullets), action recommendations (3 bullets), and a summary sentence for the manager.

Prompt to analyze 360 feedback

"Analyze this 360 feedback for [name, position]. [Paste all responses.] Generate: 1) Key strengths (3, with evidence from evaluators). 2) Development areas (3, with evidence). 3) Discrepancies between evaluators (where they significantly differ). 4) Overall feedback pattern (positive, neutral, concerning). 5) 3 recommended development actions with timeline. 6) Suggested points for the manager-employee evaluation conversation. Do NOT identify individual evaluators. Maintain anonymity."

AI-assisted calibration

Calibration is the process where managers in a department meet to compare and adjust their evaluations, ensuring that an "excellent" from one manager means the same as an "excellent" from another. Without calibration, evaluations reflect differences between managers, not between employees.

How AI assists in calibration

• Evaluation pattern detection. AI identifies managers who consistently rate higher or lower than average (lenient bias / strict bias). This is not to single out the manager, but so the group has context when discussing.
• Evidence comparison. When two employees with similar performance receive different evaluations, AI presents the evidence for both so the group can compare objectively.
• Intelligent forced distribution. If the company uses forced distribution (X% excellent, Y% good, etc.), AI can suggest the distribution based on data, preventing the decision from being purely political.
• Automatic documentation. AI generates the calibration session minutes: decisions made, justifications and adjustments. This is especially important for regulatory compliance.

From annual review to continuous feedback

The trend in performance evaluation is clear: from the annual review (once a year, long, formal) to continuous feedback (frequent, short, informal). Companies like Adobe, Microsoft, Deloitte and Netflix have eliminated or drastically transformed their annual evaluations.

The continuous feedback model

• Weekly or biweekly check-ins. Brief meeting (15-30 min) between manager and employee. It is not evaluation. It is conversation: what is going well, what do you need, how can I help.
• Real-time peer-to-peer feedback. After a project or presentation, immediate feedback from colleagues. Do not wait 6 months for the 360.
• Instant recognition. Tools that allow public recognition of a contribution the moment it occurs. Bonusly, Kudos, or simply a dedicated Slack channel.
• Quarterly evaluation (not annual). A more formal check every quarter, aligned with OKR cycles. Shorter and more frequent than the annual review.

The role of AI in continuous feedback

• Intelligent reminders. AI reminds the manager to do the weekly check-in. If it detects 3 weeks without a 1:1 meeting with a report, it sends an alert.
• Feedback suggestions. Based on OKRs and project data, AI suggests check-in topics: "Juan completed project X 2 days before the deadline. Recognize it?"
• Automatic logging. AI transcribes and summarizes check-ins (with consent), creating a history that avoids the recency effect in the quarterly evaluation.
• Recognition nudges. AI detects notable contributions (an important PR merge, a satisfied customer, a met deadline) and suggests the manager recognize them.

Bias in performance reviews

Bias in performance evaluations is systematic and well-documented. Some data:

• Women receive 2.5 times more feedback about their communication style than about their technical skills (Textio, 2024).
• Employees from ethnic minorities receive vaguer evaluations with less actionable feedback (Harvard Business Review, 2023).
• Remote employees are rated 5-10% lower than their on-site peers for the same performance (proximity bias).
• Introverted employees are rated lower in "leadership" regardless of their results.

How AI detects bias in reviews

• Language analysis. AI compares the language used in evaluations of different demographic groups. If women's evaluations use more emotional adjectives ("empathetic", "collaborative") and men's more competence adjectives ("strategic", "decisive"), there is bias.
• Score-evidence consistency. If an employee has a high score but vague written evidence, or a low score with solid evidence, AI flags the inconsistency.
• Cross-evaluator comparison. AI identifies evaluators who systematically rate employees from a certain group differently. Not as an accusation, but as data for reflection.

Prompt to review bias in evaluations

"Analyze these 10 performance evaluations [paste texts]. For each: 1) Identify if the feedback is specific and actionable or vague and generic. 2) Classify the language: results-oriented, personality-oriented, or mixed. 3) Flag if there are concrete development recommendations. Then, generate an aggregate report: are there systematic patterns in the type of language used? Are there evaluations significantly more vague than others? Are there deviations in the positive/negative feedback ratio?"

Writing evaluations with AI

Writing a good evaluation takes between 30 and 60 minutes per employee. For a manager with 10 direct reports, that is 5 to 10 hours of writing. AI can reduce that time to 5 or 10 minutes per evaluation by generating a first draft the manager reviews and personalizes.

Prompt to write a performance evaluation

"Write a performance evaluation for [name], [position], period [Q1-Q4 2026]. Data: OKRs and results: [list]. 360 feedback summary: [strengths and areas for improvement]. Notable achievements: [list]. Development areas: [list]. Generate: 1) Overall assessment paragraph (5 lines). 2) Strengths (3 bullets with concrete evidence). 3) Development areas (2 bullets with specific improvement suggestions). 4) Objectives for next period (3 proposals). Tone: professional, constructive, specific. Avoid vague language ('good overall performance') and generic phrases ('keep it up')."

Preparing difficult evaluation conversations

A manager's hardest conversation is not with the excellent employee or the terrible one. It is with the employee who is "fine but not great", who has not improved in two years, or who has a brilliant technical strength but problematic interpersonal behavior.

How AI helps prepare these conversations

• Structuring the conversation. AI generates a guide with points to cover, recommended order (start with the positive? go straight to the problem?), and key phrases for each moment.
• Anticipating reactions. Based on the employee's profile and feedback content, AI suggests possible reactions and how to handle them.
• Practicing. The manager can simulate the conversation with AI before having it with the employee. AI plays the employee role with different reactions (defensive, emotional, disinterested).

Prompt to simulate evaluation conversation

"I am going to have an evaluation conversation with an employee who: [situation description]. I want to communicate: [main message]. Simulate the conversation: you are the employee, I am the manager. The employee will probably react with [expected reaction]. After the simulation (5-6 exchanges), give me feedback on my approach: what I did well, what I could improve, and a 5-point guide for the real conversation."

Evaluation tools with AI

• Lattice: complete performance management platform with OKRs, 360, check-ins, analytics. Integrated AI for evaluation writing and bias detection. From 8 EUR/user/month.
• 15Five: focused on continuous feedback, check-ins and engagement. Less formal, more conversational. Good for companies wanting to abandon the annual review. From 4 EUR/user/month.
• BetterWorks: specialized in OKR management with advanced analytics. HRIS integrations. For medium and large companies.
• Notion/Docs + LLM: the DIY option. Evaluation template in Notion, OKRs in a table, feedback in shared documents, and an LLM to analyze and write. Works for small teams.

Ejercicio practico

Puntos clave