Christmas Eve Reflection: 4 Bets I Made in 2025 That Paid Off (and 2 That Didn't)

It is Christmas Eve and the office is empty. I have a habit of writing a year-end reflection between 24 and 31 December — partly for my own sanity, partly because every year I look back and find at least two decisions I would re-do. This year there are six. Four worked. Two did not. I am writing them down before I forget the texture of why each one was right or wrong, because the texture is what is useful. This is not a victory lap. The two that flopped cost real money and one of them cost a relationship. I am writing about them because the year-end posts I read from other founders skip the failures, and skipping the failures is what makes those posts useless. ## Bet #1 — Hiring senior over hiring fast (worked) In Q1 2025 we had budget for either three mid-level engineers or one senior engineer plus a founding designer. The conventional advice for a small consultancy is to hire mid-level — they are cheaper, you can scale capacity quickly, and the work is execution-heavy. We hired the senior plus the designer instead. Total comp was ~₹38 lakh / year for the two vs the three-mid-level option at ~₹42 lakh. Looks like a saving. Felt like a risk. The senior would either be 3x the productivity of a mid-level (paid back) or 1.2x (a bad bet). What actually happened: the senior shipped the most reliable code on the team, took the architecture decisions for two clients, and mentored the rest of us into making fewer junior mistakes. The designer ended up shaping the entire visual identity of [PenLeap](https://penleap.com) and [TalkDrill](https://talkdrill.com), which is hard to attribute a number to but I think mattered for both products' trial-to-paid conversion. The lesson I keep coming back to: cheap labour is expensive when the work is decision-heavy. Indian SMB consulting work is mostly decision-heavy. We are doing the same hiring math for 2026 — senior plus designer over three mid-level. ## Bet #2 — Going GEO-first on content, not SEO (worked, but messily) In Feb 2025 I made the call that we would write content for AI engines (ChatGPT, Perplexity, Gemini) as the primary distribution channel and treat Google rankings as a secondary outcome. The rationale was simple — our buyer profile (Indian SMB CTOs and founders) increasingly opens ChatGPT before Google for "how do I think about X" questions. If our content is what ChatGPT cites, we are in the buying conversation. What worked: the content is materially better. Writing for AI engines forces specific numbers, real client examples, and explicit citation of sources. The same posts also rank well on Google as a side effect — the [Princeton GEO study from late 2023](https://arxiv.org/abs/2311.09735) was right that statistics, citations, and quotations boost AI visibility by 30-40%, and those same signals appear to boost Google rankings. What did not work: GEO measurement is awful. There is no Search Console for ChatGPT. We track citations manually by running our key buyer queries through ChatGPT and Perplexity once a week and checking which sources get pulled. This is fragile and time-consuming. I kept hoping a vendor would ship a real GEO-tracking tool in 2025; nobody did. Net: the bet paid off in lead quality. Our 2025 inbound leads asked smarter questions and signed faster than 2024 leads. I attribute roughly 40% of that to GEO-style content. The other 60% was probably referral and word of mouth. ## Bet #3 — Going Anthropic-first on the AI side (worked) We made Claude (then 3.5, now Opus 4.5) our default model for client work in March 2025, displacing OpenAI as the primary. The bet: Anthropic's tool-use reliability and safety posture would matter more than raw benchmark scores for production agent work. The bet was right for our segment. Indian SMBs who pay us to ship agents care about reliability — a customer-facing agent that fails 1-in-50 times costs them more than a faster cheaper agent that fails 1-in-20 times. Claude's tool-call reliability has been consistently better in our production telemetry. The [pricing cut in November 2025](https://www.anthropic.com/news/claude-opus-4-5) (Opus 4.5 at $5/$25 per Mtok, down from $15/$75) made the cost objection go away. The risk we took: vendor concentration. If Anthropic had a major outage or a pricing surprise, we would feel it across most clients. We mitigated by building a routing layer that can fall back to Gemini 3 Pro or Sonnet 4.5. The fall-back has been used twice in 2025, both times during ~30-min Anthropic API hiccups. Both times the routing layer earned its keep. ## Bet #4 — Saying no to enterprise contracts (worked) In May 2025 we walked away from a ₹2.4 cr / year contract with a public-sector adjacent enterprise. The work would have been 60% of our annual revenue. The reason: the procurement cycle was 11 months, the contract had 90-day net payment terms, and the scope had no defined success criteria. I lost two nights of sleep before declining. The advice from two of my advisors split — one said "take it, the cash is real," the other said "the team will hate you for the bureaucracy and you will lose them." I went with the second advisor. We replaced the lost pipeline value with five mid-sized SMB engagements over Q3-Q4 totalling ₹1.6 cr. Smaller absolute number, much better margin, much better team morale, much faster cash. The lesson: enterprise is not always more revenue per unit of effort. For a consultancy under 30 people, an enterprise contract can absorb so much process and overhead that smaller, faster, cleaner SMB work is mathematically better. We are saying no to enterprise again in 2026 unless the cycle and terms change materially. ## Bet #5 — The WhatsApp Business API pivot (did not work) In June 2025 I pushed the team to build out a WhatsApp Business API service line. The reasoning sounded right: India is a WhatsApp-first country, every Indian SMB wants a WhatsApp chatbot, and we had built three for clients in the first half of the year. Productising it should be straightforward. It was not. The WhatsApp Business API is a moving target. Meta changed pricing twice in 2025, the template approval process is opaque and slow, and the BSP (Business Solution Provider) market consolidated in ways that made our cost structure uncompetitive. We invested ~₹14 lakh in productisation work — templates, dashboards, approval workflows — and shipped revenue of ~₹6 lakh against that investment. Net loss: ~₹8 lakh, plus 4 months of senior engineering time on the wrong thing. What I should have spotted: the unit economics were never going to work for an Indian SMB. The good WhatsApp chatbots are bundled with bigger CX platforms (Haptik, Yellow.ai), and customers who want a standalone chatbot at SMB price points are price-sensitive enough that our margin would be zero. I was anchoring on "the trend is right" instead of looking at "the unit economics are wrong." The fix: kill the service line in October 2025. Repurpose two of the engineers onto AI automation work that pays back. I should have killed it in August. ## Bet #6 — A partnership I should not have signed (did not work) In April 2025 I signed a referral partnership with a US-based consultancy. The pitch was straightforward — they would refer overseas clients who needed India-based engineering capacity, we would refer Indian clients who needed US presence. The deal was 15% revenue share both ways. What I missed in due diligence: the US firm's referral pipeline was thin. They sent us two leads in nine months, neither closed. We sent them seven leads, three closed, generating ~₹30 lakh of revenue for them — for which we received ~₹4.5 lakh in referral fees. The accounting was clean; the asymmetry was the problem. The lesson is older than 2025 and I should have known it: in a referral partnership, the side with the larger pipeline keeps most of the value. I gave away a chunk of my outbound pipeline for a tiny fraction back. The cost: ~₹30 lakh of opportunity gross, ~₹4.5 lakh of fees recovered, net loss ~₹25 lakh + the friction of working through a third party. The fix: ended the partnership in November 2025, restructured outbound for direct US client acquisition. We will close fewer US clients in Q1 2026 but keep 100% of the revenue. ## What I am writing down for 2026 ## A few smaller things that mattered in 2025

• Switched to a 4-day work week in July. Productivity went up, not down. We are keeping it for 2026.
• Started doing client invoices on the 1st of every month, no exceptions. Cash flow visibility improved by ~3 weeks.
• Moved to async standup (Slack thread, 15 min by 10 am) from daily 15-min calls. Saved roughly 45 hours/month across the team.
• Stopped doing "free first call" for inbound leads under ₹2 lakh project size. Volume of bad leads dropped 60%, conversion of good leads went up.
• Started keeping a "decisions journal" — every major call gets a one-paragraph entry with the rationale and the expected outcome. Reading it back at year end is how I wrote this post honestly.

## What this means if you are an Indian SMB founder reading this The pattern across the four bets that worked and the two that did not has one through-line: the worked bets were aligned with a structural truth about our segment (Indian SMBs want reliability, content for buyers should be GEO-shaped, senior labour pays back in decision-heavy work). The failed bets were aligned with what looked like obvious surface trends (WhatsApp is big, US partnerships look good on paper). If you are doing your own year-end review, the two questions I find most useful: (1) of every decision I made in 2025, which one would I undo if I could? and (2) of every "no" I said this year, which one would I revisit? The first surfaces the bad bets. The second surfaces the missed opportunities. 2025 had two of each, in my case. I am also setting a 2026 personal rule: every quarter I read this post again. If I am repeating any pattern from the failed bets, I stop. If you want a 30-minute year-end planning call to do the same exercise on your own decisions, I have a few slots open in the first two weeks of January. ## FAQ ### Why publish the failures publicly? Two reasons. First, the year-end posts I find useful from other founders are the ones that admit failures. Second, the founders who reach out to me after I publish failures usually have better problems to discuss than the ones who reach out after I publish successes. ### How much did the WhatsApp pivot actually cost? About ₹14 lakh in direct investment, ~₹8 lakh net loss after revenue recovery, plus ~4 months of senior engineering time on the wrong work. The opportunity cost — what those engineers could have shipped instead — is harder to estimate but probably another ₹20-30 lakh of foregone revenue. ### Why did the US referral partnership fail when the structure looked symmetric? In any referral partnership, the side with the larger outbound pipeline transfers value to the side with the smaller pipeline. We had a stronger Indian SMB pipeline than the US firm had US enterprise. The 15%/15% structure was symmetric on paper but asymmetric in practice. Always model the absolute revenue flow, not the percentage. ### Should consultancies under 30 people take enterprise contracts at all? Sometimes — if the procurement cycle is under 4 months, payment terms are 30 days or better, and there is a defined success criterion. Otherwise the overhead of running enterprise process inside a small team will burn out your senior people and starve your other clients. ### Is the 4-day work week sustainable for a consultancy? So far yes, in our experience. The trick is that we kept the same project commitments and trusted the team to compress. Productivity per hour went up — meetings shrank, deep-work time grew. We will revisit at the 12-month mark in mid-2026. ### What is the "decisions journal" you mentioned? A simple Notion page where every major call (hire, fire, sign a contract, decline a contract, change a service line) gets a one-paragraph entry — the date, the decision, the rationale at the time, the expected outcome. Reading it back six months later is the most honest way I have found to assess my own judgment. ### Will you publish this kind of reflection annually? Yes. The Christmas Eve format works for me. Expect the same structure at the end of 2026 — a list of bets that worked and bets that did not, with real numbers and the lessons.